Final Devoicing in the Phonology of German [Reprint 2012 ed.] 9783110966060, 9783484303362

Citation preview

Linguistische Arbeiten

336

Herausgegeben von Hans Altmann, Peter Blumenthal, Herbert E. Brekle, Gerhard Heibig, Hans Jürgen Heringer, Heinz Vater und Richard Wiese

Wiebke Brockhaus

Final De voicing in the Phonology of German

Max Niemeyer Verlag Tübingen 1995

"Für meine Eltern, die mich bei meiner Arbeit immer unterstützt haben. "

Die Deutsche Bibliothek - CIP-Einheitsaufnahme Brockhaus, Wiebke: Final devoicing in the phonology of German / Wiebke Brockhaus. Tübingen : Niemeyer, 1995 (Linguistische Arbeiten ; 337) NE: GT ISBN 3 - 4 8 4 - 3 0 3 ^ - 0

/ é

-

ISSN 0344-6727

© Max Niemeyer Verlag GmbH & Co.KG, Tübingen 1995 Das Werk einschließlich aller seiner Teile ist urheberrechtlich geschützt. Jede Verwertung außerhalb der engen Grenzen des Urheberrechtsgesetzes ist ohne Zustimmung des Verlages unzulässig und strafbar. Das gilt insbesondere für Vervielfältigungen, Übersetzungen, Mikroverfilmungen und die Einspeicherung und Verarbeitung in elektronischen Systemen. Printed in Germany. Druck: Weihert-Druck GmbH, Darmstadt Einband: Hugo Nädele, Nehren

Table of contents

Abbreviations and Notational Conventions

ix

Acknowledgments

xi

Introduction and Overview

1

1. The 'Facts' of Final Obstruent Devoicing 1.1 Introduction 1.2 FOD in German 1.2.1 Historical development 1.2.2 The application of FOD in present-day German Hochlautung 1.2.2.1 Introduction 1.2.2.2 Voicing alternations 1.2.2.3 Distributional constraints affecting voicing 1.3 Conclusion

3 3 6 6 6 6 11 22 23

2. 25 Years of FOD: Earlier Approaches from 1968 to 1993 2.1 Introduction 2.2 What is FOD? 2.2.1 The textbook solution: [- son] -»• [- voice] 2.2.2 The textbook solution in a different guise: [- son] [+ tense] 2.2.3 Devoicing as delinking/deletion 2.2.4 Constraints on the licensing of a privative [voice] feature 2.2.5 Conclusion 2.3 The link between the 'what' and the 'where' 2.4 Where does FOD occur? 2.4.1 Introduction 2.4.2 FOD applies preceding a #-boundary 2.4.2.1 Vennemann 1968 2.4.2.1.1 The analysis 2.4.2.1.2 Problems with Vennemann's 1968 analysis 2.4.2.1.2.1 Using a #-boundary to represent a syllable-boundary 2.4.2.1.2.2 Deleting a single internal #-boundary before a vowel 2.4.2.1.2.3 Dealing with variable items: 'unpronounceable' clusters 2.4.2.1.2.4 Summary 2.4.2.2 Kloeke 1982 (and Wurzel 1970) 2.4.2.2.1 The analysis 2.4.2.2.2 Problems with Kloeke's analysis 2.4.3 FOD applies immediately before a syllable boundary

24 24 25 25 27 31 33 34 34 37 37 37 37 37 41 41 44 46 49 49 49 51 54

2.4.3.1 Introduction 2.4.3.2 Wiese 1988 and Hall 1992b: the analysis 2.4.3.3 Wiese 1988 and Hall 1992b: some potential problems 2.4.4 FOD applies to segments in a coda 2.4.4.1 Introduction 2.4.4.2 Rubach 1990: the analysis 2.4.4.3 Some problems with Rubach's analysis 2.4.5 The feature [voice] is licensed only in a syllable onset: Lombardi's (1991) analysis and its problems 2.5 Conclusion

54 54 65 67 67 68 72 85 88

3. What is Final Obstruent Devoicing? A Government Phonology Approach 3.1 Introduction 3.2 Phonological licensing, government, 'ground rules' and some principles of grammar 3.2.1 Phonological licensing and government 3.2.2 'Ground rules' and some principles of grammar 3.3 Phonological elements and segmental representations in GP 3.3.1 Phonological elements in GP 3.3.2 Segmental representations of some German obstruents 3.3.3 Segmental representations of some German sonorants 3.4 What are the changes involved in FOD? 3.4.1 What is 'voicing'? 3.4.1.1 Some general cues to the voicing contrast 3.4.1.2 F0 perturbations and the voicing contrast 3.4.2 What are the options for FOD made available by GP? 3.4.3 Final obstruent devoicing as loss of lexical L 3.4.3.1 A typical weakening process? 3.4.3.2 What are the implications of losing L? 3.4.3.2.1 FOD is not a neutralisation process in the sense of Kiparsky 1976 3.4.3.2.2 Syllable-final devoicing is not a dynamic phonological process . . 3.4.3.2.3 Some devoiced obstruents can govern only homorganic nasals . . . 3.5 Conclusion

89 89

129 139 149 156

4. Where Does Final Obstruent Devoicing Occur? A Government Phonology Approach 4.1 Introduction 4.2 Straightforward FOD environments 4.2.1 Word-final and term-final in compounds 4.2.2 Before certain suffixes 4.3 FOD asan instance of autosegmental licensing 4.4 FOD in Vennemann's variable items

158 158 158 158 165 171 185

90 90 99 104 104 107 112 116 116 116 119 123 127 127 129

vii

4.4.1 Introduction 4.4.2 FOD preceding domain-internal licensed empty nuclei 4.4.2.1 Impossible alternative structures 4.4.2.2 Proper government and the Empty Category Principle 4.4.2.3 An alternative solution: differences in morphological structure between Hochlautung and NSG 4.4.2.4 More on problematic clusters 4.4.2.5 Summary and interonset licensing 4.4.2.6 Schwa-ful versus schwa-less forms 4.5 Some problems with Rubach's analysis revisited 4.5.1 Introduction 4.5.2 Native speaker judgments on the location of syllable boundaries . . . . 4.5.3 The special properties of /r/ 4.6 Apparent failure of FOD to apply word-finally 4.7 Conclusion

185 188 188 196 198 201 206 211 226 226 226 228 233 236

5. Final Obstruent Devoicing and Neutralisation 5.1 Introduction 5.2 Neutralisation and the traditional place of FOD 5.2.1 Two views of neutralisation: Trubetzkoy and Kiparsky 5.2.2 Earlier analyses of FOD and their implications for FOD as a neutralisation process 5.3 The neutralisation debate in the experimental literature 5.3.1 The studies and their findings 5.3.2 The relationship between production and perception 5.3.3 The implementation of FOD 5.3.4 Experimental studies and phonological neutralisation 5.3.5 Summary 5.4 The GP analysis 5.4.1 The proposals 5.4.2 Phonological evidence 5.4.2.1 Spirantisation of velars: two analyses 5.4.2.2 Why do only velars spirantise? 5.4.3 Experimental evidence 5.4.4 The general implications for earlier problems 5.5 Conclusion

238 238 238 238 241 242 242 245 246 248 249 250 250 251 251 257 259 260 261

References

263

Abbreviations and Notational Conventions

1. Abbreviations acc. adj. C coll. dat. DFA DP ECP fam. fem. FOD gen. GP ind. imper. mase. MHG MLG MSC Ν NHG nom. NSG OCP 0 OHG pl. pres. pol. R RP RVA SD sg· so. SSA SSC

accusative adjective consonant colloquial dative Dorsal Fricative Assimilation Dependency Phonology Empty Category Principle familiar form feminine final obstruent devoicing genitive Government Phonology indicative imperative masculine Middle High German Middle Low German Morpheme Structure Condition nucleus New High German nominative Northern Standard German Obligatory Contour Principle onset Old High German plural present polite form rhyme Received Pronunciation regressive voicing assimilation structural description singular someone Syllable Structure Algorithm Syllable Structure Condition

χ SSG UG V VOT

Sonority Sequencing Generalisation Universal Grammar vowel voice onset time

2. Transcription Throughout this book, I have made transcriptions no narrower than necessary. Aspiration, for example, is almost never indicated, although frequently present. The phonetic symbols used have their IPA values, with the following exception: [r] = consonantal /r/-realisation (i.e. [Κ, explanation)

[Ç],

[R], or [r]; see

1.2.2.1

for a more detailed

Slashes (/ f) indicate that a form is underlying or otherwise potentially remote (without making any claims about the status of a segment in terms of phonemic classification), while square brackets ([ ]) are associated with surface realisations. Transcriptions quoted from other authors have been adjusted to fit into the system used in this book.

Acknowledgments

My greatest intellectual debt in writing this book is undoubtedly to John Harris. He supervised me in my postgraduate days at UCL and, during that period, he shared many of his ideas with me and was always a source of sound advice and intellectual as well as practical support. I have no doubt that much of what is worthwhile about the present work has its roots in what I learnt from him. He also read an earlier draft of chapters 3 and 4 and provided me with numerous helpful comments. I am greatly indebted to Richard Wiese, who read the entire first draft of the manuscript with great care and made a considerable number of insightful observations and suggestions for changes. I would like to thank Stamatoula Pagoni, Edmund Gussmann, Gienek Cyran, Robert Foltin and Péter Szigetvári for valuable comments on an earlier draft of chapters 3 and 4. I am also grateful to Geoff Williams for his perceptive observations on chapter 3. Although I considered all their queries and suggestions very carefully, I decided that I knew better on a few occasions - and will probably be proved wrong in the end. None of those who have read drafts agree with everything I say here, and they certainly should not be held responsible for any inadequacies that remain. Special thanks go to Monik Charette, who was always willing to spend time discussing issues arising from my analysis and who supplied some useful intuitions on French. I am indebted to Susan Barry, John Coleman, Jacques Durand, Ken Lodge, Geoff Lindsey, Max Wheeler and Shohei Yoshida, who have all helped in some way, be it by commenting on my work or by providing me with useful information. Michael Ingleby, Carl Chalfont and my colleagues at the University College of Ripon and York St. John have made a particular contribution by giving me the space I needed to finish the book. Thanks are also due to my patient informants, especially Gertrud and Herbert Brockhaus, and Use and Werner Wolf. Finally, my sincerest thanks go to Jenny Sutton, who has accompanied this project from the very beginning and who was always willing to provide practical and moral support. I am particularly grateful to her for generously giving of her time and her considerable skill with WordPerfect's Draw facility to create the diagrams.

Introduction and Overview

This book deals with the phonological event of final devoicing in a theoretical framework based on principles and parameters rather than rules. It refers to data coming almost exclusively from German (native and non-native items). The first chapter presents the 'raw facts' of final devoicing and voicing restrictions on obstruent clusters. Its purpose is to provide an outline of the sort of alternations and distributional restrictions to be accounted for in the remainder of the book. It also constitutes a first attempt at imposing some kind of order on the data and establishing what the typical cases are on which any analysis must focus, at least to begin with. Previous treatments of final devoicing in German are discussed and evaluated in the second chapter. Many textbooks are still dominated by the SPE-type rule first proposed by Vennemann in 1968 ([- son] [- voice]/ _ #). However, contrary to the impression created by such publications, a great deal of progress has actually been made in detailed studies of final devoicing since then. As far as the 'what' of this phonological event is concerned, a development can be traced from an arbitrary rewrite rule referring to a binary feature [± voice] (e.g. Vennemann 1968) to constraints on the licensing of a privative feature [voice] (e.g. Lombardi 1991). The 'where' of final devoicing, which has always received more attention than the 'what', has moved on from reference to #boundaries (e.g. Vennemann 1968, Kloeke 1982a) to analyses which crucially involve the syllable, be it a syllable boundary (e.g. Wiese 1988, Hall 1992b) or a syllable coda (e.g. Rubach 1990). Most recently, an approach has been proposed which can be described as the converse of earlier syllable-based accounts. It makes reference only to syllable onsets and argues that final devoicing applies elsewhere (Lombardi 1991). In the course of the chapter it becomes clear, however, that there are problems associated with all available accounts of final devoicing. Chapters 3 and 4 provide an analysis of final devoicing in German couched in the framework of Government Phonology (e.g. Kaye el al. 1985, 1990), a phonological theory operating with principles and parameters, not rules. Some of the central tenets of this theory are introduced at the beginning of chapter 3, but additional concepts are explained as they become relevant to the discussion of final devoicing. Apart from the rejection of rules, Government Phonology also deviates from much of mainstream phonology in that the primitives are not binary or even privative features, but 'elements'. These are cognitive entities which map onto acoustic (and articulatory) correlates. They are pronounceable in isolation or in combination with each other at all levels of derivation, so that underspecification has no role to play in Government Phonology. The third chapter addresses the question of what final devoicing is. The conclusion drawn is that final devoicing is a phonological weakening process involving the loss of the laryngeal element L (which represents voicing in obstruents). Arguments presented in favour of this kind of approach are both theory-internal and external. The theoryinternal arguments overlap with the theory-external ones where consideration is given to the environment in which final devoicing occurs. Whatever theoretical interpretation may

2

be applied to it, this environment (word-finally, syllable-finally) can be shown to be a 'weak' environment, that is, an environment in which segments are particularly prone to lenition. In Government Phonology, the loss of an element (e.g. L) is interpreted as segmental weakening. Distributional restrictions, as opposed to voicing alternations, can be handled in a similar way. I show that many so-called syllable-final obstruents occupy what, in Government Phonology, must be interpreted as rhymal adjunct positions. Rhymal adjuncts cannot accommodate L in German - or, it seems, in any other language. However, in languages such as English, these positions may share the laryngeal element of the following onset position, so that certain voicing assimilation effects are created (e.g. e[ks]it or efgzjit but not *e[kz]it or *e[gs]it). In German, by contrast, no such sharing can occur, so that the first obstruent in a consonant cluster will always be voiceless. The discussion of the 'what' is followed (in the fourth chapter) by an investigation of the trigger of final devoicing. Government Phonology recognises empty nuclear positions. Broadly speaking, these may be licensed to be empty by parameter setting (domain-final positions only) or through the establishment of licensing relations with other, unlicensed nuclear positions. I adapt claims about final devoicing made in the third chapter to accommodate insights gained in the fourth and argue that final devoicing is triggered where the element L cannot be autosegmentally licensed (Goldsmith 1989, 1990). In other words, final devoicing is not a matter of delinking L but of withdrawing its licence. The conditions under which it is impossible for L to be licensed can be identified with reference to the notion of Licensing Inheritance (Harris 1992). This predicts that an onset position immediately preceding a domain-final empty nuclear position will exhibit reduced autosegmental licensing potential. I show that word-final obstruents subject to devoicing occupy precisely such positions. This part of my analysis also accounts for final devoicing before suffixes such as -lich, -bar, -los, -t, -s etc. For speakers of Northern Standard German, unlike speakers of the prestige variety (Hochlautung), final devoicing can also apply morpheme-internally. Here, it can be demonstrated that licensed empty nuclear positions are, again, implicated. I discuss how these nuclear positions receive their licensing and how certain schwa/zero alternations should be dealt with. From this discussion it emerges that German has no lexical schwas and that all schwas can be interpreted as manifestations of unlicensed empty nuclei. The final chapter takes a new look at the concept of neutralisation in the context of final devoicing in German. It briefly highlights two different views of neutralisation, those of Trubetzkoy (1939 [1969]) and Kiparsky (1976), and discusses some of the experimental work which has examined final devoicing as a putative process of neutralisation over the past decade or so. It points out that some of the findings of this work pose problems for the phonological theories which first inspired the experimental research activity in this area. The Government Phonology analysis of final devoicing presented in chapters 3 and 4 is then confronted with these problems. I demonstrate that they can be resolved, given this particular approach.

1. The Tacts' of Final Obstruent Devoicing

1.1 Introduction Some loss of voicing in final obstruents (most commonly in word-final position, especially when the word is utterance-final) is attested in a vast number of languages. English is probably the most well-known language to exhibit this apparently physiologically conditioned property. Phonologically speaking, however, this loss of voicing is not significant, as the affected segment is still perceived and interpreted as the voiced member of a pair of obstruents. It is for this reason that minimal pairs such as nib - nip, bed - bet and lag - lack can exist in certain dialects of English (e.g. RP, also known as Southern British Standard or BBC English; see Wells 1982: 117). By contrast, there is a sizeable group of other languages all over the world where the loss of voicing is magnified to such an extent that native speakers typically perceive a final obstruent as phonologically voiceless, i.e. as the voiceless member of a pair of obstruents. It is not possible for, say, German hearers to reliably distinguish the noun Bund ([bunt], 'league') from the adjective bunt ([bunt], 'colourful') when the two are spoken in isolation. This apparent absence of phonological opposition, however, cannot be observed if the relevant segment is in an inflected form, such as Bunde (['bundg], league', dat. sg.) vs. bunte (['bunta], 'colourful', fem. nom. sg.). The phonological event which consists in this loss of voicing is most frequently referred to as final devoicing1 and, occasionally, as obstruent devoicing1 or terminal devoicing,3 German-language publications4 know it as Auslautverhärtung, while I shall refer to it as final obstruent devoicing (or FOD for short) in this book, to distinguish it from other alleged final devoicing processes, such as so-called final sonorant devoicing in Polish, Icelandic, Russian, Danish (optional) or Angas,5 for example. In what follows, voicing is to be understood in its broad traditional sense.6 Phonologists working in a variety of different theoretical frameworks have taken an interest in final obstruent devoicing7 - described by Giegerich (1989: 51) as 'this most

1

2

3 4

5

6 7

For example, by Vennemann (1968:161,1972:12), Hyman (1975: 71 f.), Kenstowicz & Kisseberth (1979: 337), Kaye (1989: 59), Rubach (1990) and Hall (1993). For example, by Katamba(1989: 167), Durand (1990: 182), Rischei (1991), Giegerich (1992) and Roca (1994: 142). For example, by King (1969: 47), Escure (1975: 33) and Dinnsen (1980). E.g. Wurzel 1970, Kloeke 1982a, b, Meinhold & Stock 1982, Vennemann 1982a, Wiese 1988,1991 and Yu 1992a, b. See, for example, Gussmann 1992, Escure 1975 (pp. 165f ), Barry 1989, Halle & Clements 1983 (p. 45) and Rischel 1991.1 will return to this process, albeit briefly, in 3.4.2. The traditional use of the term and its implications will be discussed in detail in 3.4.1. See the following chapters for detailed references.

4

popular of German phonological rules' - for a number of reasons, some of which will become apparent in chapter 2, where I discuss several analyses put forward over the past twenty-five years or so. Apart from these, three of the most obvious reasons for FOD's popularity, as far as I can see, are the following. Firstly, final obstruent devoicing is found in one of the most thoroughly-studied languages of the world, German. Secondly, it is a phonological event which can be described as a '"soft" universal - a configuration of data that is found in many languages' (Kenstowicz 1994: 495). It crops up all over the world, not just within the Germanic group of Indo-European languages (e.g. German and Dutch) but also in Balto-Slavic languages, such as Russian, Polish and Czech, as well as Romance languages like Catalan. In fact, it is amply attested outside the Indo-European family. Unrelated languages, such as Wolof (West Atlantic; Senegal, Gambia), Ojibwa (Macro-Algonquian; central and eastern Canada, northern central United States) and Turkish (Altaic; Turkey, Bulgaria), for example,8 also exhibit FOD. Thirdly, there seems to be something 'deep-seated' about it, which makes researchers feel that an understanding of FOD may be a particularly useful key to phonology. This deep-seatedness manifests itself in the fact that languages apply FOD to foreign loan words with the same regularity as they do to native words and that individual speakers with few exceptions - carry FOD from their mother tongue over into other languages they may be learning or may have learnt as adults. In fact, even speakers of so-called CV languages, with no final obstruents, apply FOD in other languages. This, of course, is only noticeable where that other language does not itself exhibit FOD. For example, the application of FOD is one of the hallmarks of a German accent of English. Another aspect of the 'deep-seatedness' of FOD is that, during the process of first-language acquisition, children tend to go through a phase of applying FOD (usually when obstruents are first acquired), even in languages where adult speakers don't (Stampe 1969). The acquisition of English as a first language, for example, tends to involve a phase of so-called context-sensitive voicing, where word-initial obstruents are exclusively voiced and word-final obstruents exclusively voiceless. Context-sensitive voicing usually begins to disappear by 2;6, although 'some children continue to use voiceless [...] obstruents in word-final position for some time after 3;0 for all types of obstruents' (Grunwell 1987: 226). Finally, the application of FOD can even be observed in adult native speakers of non-FOD languages such as English - when these speakers are in a state of inebriation. It may not be the case that consistent across-the-board application of FOD can be observed, but a definite tendency towards final devoicing has been reported in drunken speakers by Lester & Skousen (1974). The purpose of this chapter is to present the 'facts' of FOD. As pointed out by Popper (see Magee 1982: chapter 2), however, it is not actually possible to present pure facts without reference to theory, or, as Magee (1982: 33) puts it, 'observation as such cannot be prior to theory as such', hence the quotation marks around the word facts. All I can do 8

See, for example, Maddieson 1984 (pp. 174ff ), Katzner 1986 or Crystal 1987 (pp. 436ff.) for details of the classification of the world's languages.

5

is to offer those facts which are relevant to the phenomenon on the broadest theoretical basis possible (i.e. a combination of all analyses of FOD I am aware of). It is worth noting, though, that even a broad overview of the 'facts' needs to have some kind of order imposed on it. The order I intend to impose is one which begins to whittle down large chunks of data and attempts to pick out the clear-cut, typical cases of FOD. So, as well as lists of data, this chapter contains critical discussion of unusual cases and an attempt to distil the essence of FOD in German. My own analysis of FOD (to be developed in chapters 3 and 4) takes this essence as its point of departure and is then extended to account for some of the more unusual cases, too. I will present data sets of two kinds, those which illustrate the sort of alternations FOD gives rise to and those which illustrate the distributional restrictions on voicing which are usually discussed in the context of FOD (e.g. in Vennemann 1978 or Giegerich 1989). Both illustrate the point that FOD is not necessarily restricted to word-final position. From cross-language studies such as Mascaré 1987a or Lombardi 1991 and even the most superficial inspection of other FOD languages, such as Dutch,9 Russian, Catalan or Polish, for example, it is clear that there is some variation between languages regarding die environments where FOD occurs. No doubt, the nature of this variation can reveal a great deal about the special properties of FOD, as can looking at FOD in the context of voicing assimilation processes (which are practically non-existent in German). Ideally, then, a study of FOD should be a cross-language study, pooling data from a range of languages in the manner of Mascaré 1987a or Lombardi 1991. This is, in fact, what I set out to do when first embarking on the research for this book. However, my investigation of FOD in German (my mother-tongue), couched in the framework of Government Phonology (see 3.1 for references), raised a considerable number of issues which needed to be followed up and which are discussed in chapters 3 to 5. It seems to me that gaining an understanding of FOD in other languages would require equally thorough investigation and reanalysis of familiar data in the GP framework. This book may then be seen as the first stage in a larger cross-language study of FOD (the completion of which, if it turns out to be desirable, is bound to involve a number of researchers, not just me). As Vennemann (1983: 7), puts it, 'given the limitations of the human existence, all theories about language have to be partial theories [...]. We create partial theories with the awareness they are partial, that they have to fit together with other partial theories [..].' Although I would not claim that what follows amounts to a theory in any strict sense, I hope that it will turn out to fit together with work done by others and make a contribution to the quest for a better understanding of UG by raising issues which are relevant universally. Let me now turn to the 'facts' of FOD in German.

9

See, for example, Booij & Rubach 1987 or Berendsen 1987 for Dutch, Hayes 1984 or Kiparsky 1985 for Russian, Wheeler 1979,1983,1986 or Mascaré 1987b for Catalan and Rubach 1982 (pp. 120ff), 1984 (pp. 206f), Kenstowicz & Kisseberth 1979 (pp. 418ff ), Jassem & Richter 1989, Slowiaczek & Dinnsen 1985, Booij & Rubach 1987 and Gussmann 1992 for Polish.

6

1.2 FOD in German 1.2.1 Historical development Detailed and reliable information on the emergence of phonologically significant final devoicing is almost impossible to come by. Most historical works on German10 mention this process only in passing, while focusing on the more dramatic changes affecting the language at that time. Researchers whose main interest is final devoicing, on the other hand, do not usually concern themselves with its diachronic status. The most notable exception is probably Parker (1981), who deals with the issue of why final devoicing should have arisen historically. Detailed discussion of his approach would, however, be beyond the scope of this book, so I shall have nothing to say about why final devoicing should exist. For the present section, I would like to do no more than locate FOD historically. Final devoicing found its way into most German dialects (there was no unified and widely-spoken standard pronunciation at that time) as one of the numerous changes which took place during the transition from Old High German (c. AD 750-1050) to Middle High German (c. AD 1050-1350; see Walshe 1974: 2). High German (as in the terms Old High German and Middle High German), in this context, does not refer to a non-regional standard variety (as it does today) but to those mainly Upper German dialects (i.e. Alemannic and Bavarian-Austrian) which were spoken by the authors of the literary manuscripts from which most of our knowledge of the language ofthat period is derived.11 What is clear from these sources is that final b, d, g became p, t, k, as indicated by MHG spellings such as lip ([li:p], "body'; gen. sg. libes [liibas]) and tac ([tak] 'day'; gen. sg. tages ['tagas]) (see König 1978: 73). There is less agreement, however, on whetherfricativesbecame subject to this process at the same time, although some researchers suggest that this was, in fact, the case. Priebsch & Collinson (1958: 119) quote Ao/('court'; gen. hoves) as an example of final devoicing of fricatives. This is how far the rather meagre, but relatively undisputed, history goes. I now turn to the more richly documented present.

1.2.2 The application of FOD in present-day German Hochlautung 1.2.2.1 Introduction Before dealing with the details of FOD in German, let me first of all identify the kind of German that is to be investigated. This study is concerned mainly with the variety of German which corresponds roughly to what Wells (1982: 279ff.) labels mainstream RP 10

11

E g. Priebsch & Collinson 1958, Bach 1961, Chambers & Wilkie 1970, Penzl 1975, König 1978 and Szulc 1987. See Walshe 1974 (pp. 2ff.) for further details of this classification.

7

or, perhaps, even near-RP for British English. In the most recent edition of the authoritative DUDEN Aussprachewörterbuch12 (Mangold et al. 1990), this variety is referred to as Standardlautung. It is free from regional characteristics and, therefore, suitable for use by radio and television announcers. It is also the variety taught to foreign learners. I will refer to it as Hochlautung (rather than Standardlautung), since I believe that most readers will be more familiar with this traditional term. Although I am mainly concerned with German Hochlautung in this book, the actual dialect to be described (my own) differs from Mangold et. al.'s interpretation of Hochlautung in two respects. The first concerns the vowel inventory and the second the environments in which orthographic r is realised vocalically. These differences constitute a slight northern influence on my pronunciation and are minor enough for the dialect still to be describable as Hochlautung. The vowel inventory of my version of Hochlautung lacks [ε:], which is the long realisation of orthographic ä, for example in a word such as Käse ([lceiza], 'cheese'). Speakers of this dialect use [e:] instead of [ε:], so their pronunciation oí Käse is [lceiza]. This is reflected in the transcriptions contained in the remainder of the book. It is not immediately relevant to my discussion of FOD, but still worth noting. The second difference between the dialect from which the vast majority of the data in this book is drawn and Mangold et al.'s interpretation of Hochlautung - as already mentioned - concerns the question of where orthographic r is realised vocalically. Before I can describe the details of this, I need to say a brief word about the consonantal realisations of r. Modifying slightly the relevant observations made by Mangold et. al. (1990: 45f.), it seems fair to say that there are four different types of consonantal r-realisation in Hochlautung, viz. an alveolar tap [r], a uvular trill [R], a uvular fricative [Κ] and a uvular approximant [ç], The apical realisation is favoured by southern speakers, while the uvular ones are predominant in northern areas and are also adopted by trained speakers. It appears that the latter are more or less in complementary distribution, with aerodynamic factors playing a crucial role in determining which uvular realisation is produced.13 Given this variation of consonantal r-realisations, I shall follow the practice adopted by Mangold et al. (op. cit.) and represent consonantal /r/ as [r], which is intended to represent not an alveolar trill, but any of the consonantal r-realisations described above. How [r] is to be interpreted phonologically will be discussed in 4.5.3. According to Mangold et al. (1990: 45f.), orthographic r is realised consonantally before long and short vowels as well as diphthongs, and after short vowels (where r itself may be either word-final or precede another consonant). So, according to the DUDEN

12

13

The DUDEN is renowned in Germany as the authoritative reference publication on the German language. Its most important manifestations today are the 10-volume DUDEN, where each volume covers a particular area such as grammar, etymology, pronunciation or orthography, and the sixvolume Großes Wörterbuch der deutschen Sprache, which combines information on all of the above aspects (as well as usage) in an alphabetically arranged dictionary of over 500 000 entries. See Hildebrandt & Hildebrandt 1965 for discussion.

8

pronouncing dictionary, we find prevocalic [r] in words such as Raum ([raum], 'room, space') and Karre ([kara], "barrow') and postvocalic [r] in, for example, dürr ([dyr], 'arid') and Wort ([vort], 'word'). After long vowels, orthographic r is realised as an unrounded central vowel between open-mid and open (according to the 1989 IP A classification). This is true both word-finally (e.g. in Bier [bi:ç], "beer') and before a consonant (e.g. in Pferd [pfe:çt], 'horse'). Unlike Hochlautung (or, at least, Mangold et al.'s interpretation of it), the dialect I describe in this book has /"-vocalisation after both long and short vowels, regardless of what follows. So, consonantal [r] is restricted to the sort of prevocalic positions just described, while words such as dürr and Wort are pronounced [dYç] and [voçt] rather than [dYr] and [vort], (Incidentally, according to Vater (1992: 110), r-vocalisation after short vowels is now typical of'standard' Hochlautung, not just of the marginally northern variety featured in my description.) This fact will be reflected in transcriptions (and analysis) throughout the book. Apart from these two genuine dialect differences, there is a third area in which my transcriptions systematically differfromthose in Mangold et al. 1990. This concerns the pronunciation of tautosyllabic (often post-obstruent) /a/ + Imi, Ini or IV. Here, schwa (or even [ε], see Mangold et al. 1990: 58f.) is present only in particularly slow, careful or emphatic speech, that is, in what Vennemann (1982a: 276) refers to as 'over-articulation' and what Höhle & Vater (1978: 173) perceive as 'very distinct pronunciation'. This kind of articulation is not uncommon in citation forms (Giegerich 1989: 50), but it does not tend to be used in connected speech under normal circumstances. What happens instead is that the schwa is not realised and the sonorant becomes the syllable peak. This results in realisations such as [?alqi] (allem, 'all', mase. dat. sg.), [fa:drj] (Faden, 'thread') and [Jtru:dJ] (Strudel, 'swirl').14 To the extent that the behaviour of schwa and the syllabic vs. non-syllabic status of sonorant consonants are of importance for my discussion of FOD, I will take them into account. In fact, chapter 4 contains detailed discussion of the properties of schwa and of true schwa/zero alternations. These are distinct from the cases just cited in that schwa genuinely 'disappears' instead of manifesting itself through the syllabicity of a following sonorant. The pair ['ro:dal/ro:d}] - ['ro:dle] (Rodel - Rodler, 'toboggan - tobogganist') would be an example of such an alternation. However, the issue of schwa + sonorants vs. syllabic sonorant lies largely outside the main focus of the present study, so that, for greater simplicity, I will spell out the schwa in forms such as [?alam], ffaidan] and [Jtru:dal], The main exception to this strategy involves /a/ + Irl sequences which appear as they are normally realised in connected speech, i.e. as [B]. To conclude this discussion, let me clarify the status of [e] vs. [ç]. The former is the realisation of an underlying /a/ + Irl sequence (orthographic -er),while the latter represents vocalised /r/ (orthographic r). According to Mangold et al. (op. cit.), [Β] and [ç] have the same place of articulation but differ from each other in that the former is

14

See e.g. Höhle & Vater 1978 and Wiese 1988 (pp. 168ff.) for discussion of the role of schwa in the formation of syllabic sonorants.

9

syllabic, while the latter is not. How this is reflected in segmental representations (or, more accurately, prosodie structure) will be discussed in 4.5.3. In later chapters, varieties other than Hochlautung, especially what Vennemann (1972: 12) labels Standard German, Northern pronunciation, will also be referred to. I shall adopt the term Northern Standard German (NSG) for this variety. To begin the discussion of FOD, we need to establish which segments are affected by it. Obviously, these must be obstruents. What exactly the members of the German obstruent inventory are and exactly which of these participate in FOD is, however, less clear. Consider the following chart of'underlying [+ consonantal] segments'fromHall 1992b (p. 21).

(1) bilabial stops

affricates

labiodental

dental

palatoalveolar

palatal

velar

Ρ

t

k

b

d

g

pf

ts

uvular

glottal

tJ (d 3 )

fricatives

nasals liquids

m

f

s

J

ν

ζ

(3)

h

Ç

η 1

R

Several of the implicit claims in (1) are controversial (e.g. underlying /ç/; see, for example, Macfarland & Pierrehumbert 1991, Hall 1989b, 1992b (chapter 5), and Iverson & Salmons 1992 for recent discussions and useful references), but I shall put most of these to one side and focus on those segments which appear to be members of a voiced/voiceless pair. There are seven such pairs, two of which involve a bracketed voiced member, viz. ( % ) and (Jzf). Hall (1992b: 15) justifies these brackets by pointing to the fact that both segments occur only in recent loan words. His view that these segments have some kind of special status is, indeed, shared by a number of phonologists, some of whom15 even take the more drastic step of excluding IATJ andtyfromtheir respective consonant charts

15

These authors include Lass (1984: 94,152), Meinhold & Stock (1982: 121) and Fox (1990: 30).

10 altogether. Other authors list /$/, but make no reference to /dj/. 16 Given that the general consensus appears to disfavour /dj/ and the fact that there are no examples of voicing alternations involving /tj/ and ld¡l, I shall ignore M3/ in the remainder of the discussion of FOD. The fricative I3I, by contrast, appears to merit further discussion. After all, several authors do consider it to be in some way part of the German consonant inventory and, more importantly, for the present purposes, /// and I3I seem to alternate. As Richard Wiese (p.c.; see also Wiese 1990) has pointed out to me, an apparent alternation pair orange [o'raqi] 'orange' (colour) and orange [o'raq3a] 'orange' (colour), fem. nom. sg., exists in German Hochlautung. This is actually denied in Mangold et al., where only [o"r^a] and [otaqja] are listed. My own intuitions match Wiese's, though, so that I am reluctant to reject his observation out of hand. However, this alternation is, at best, atypical, for the following reasons. Firstly, as observed by Hall (1992b: 15), the voiced member of the pair, /j/, unlike the voiceless member, ///, occurs only in loan words (mainly from French; see Werner 1974: 47), which are still readily identifiable as such (see discussion on p. 13 below). In colloquial speech, especially that of uneducated speakers not familiar with French, there is a tendency for [3] to be replaced by [J] (see e.g. Mangold et al. 1990: 58). Many speakers have [ga'ra:J] rather than [gatada] (Garage, 'garage') and [lcu'ra:J] instead of [ku'raga] (Courage, 'courage'). This 'normalisation' suggests that ΙζΙ is a marginal member of the German consonant inventory. It reaches even more dramatic proportions for some speakers in the context of the colour adjective beige ([be:/]), which is also borrowed from French. The feminine singular and nominative plural forms, although identical with the nominative singular form in the spelling (like orange), are realised with schwa, so that the palato-alveolar fricative is no longer devoiced (['be:3a]) in Hochlautung. However, some speakers (especially in the South) do not attempt to imitate the French pronunciation, but have ['baiga] instead, which provides additional support for my claims about the marginal status of [3]/[J] alternations. A further argument in favour of saying that I5I is a marginal member of the German consonant inventory is that, unlike all other non-glottal, non-affricate segments listed in (1), it never occurs adjacent to another consonant. The only apparent consonant which may be found next to is Irl. But the consonantal realisation is ruled out in these environments, so that we hear [53] rather than [13]. As we shall see in chapter 3, it is perfectly straightforward to account for the fe]/[J] alternation in GP, but, there can be no doubt about its marginal status. Because of this and the very limited amount of relevant data, I propose to exclude ty from consideration, so that we are left with five pairs of obstruents which have a voiced and a voiceless member each, viz. the three stop pairs /p, b/, It, d/ and Ik, g/ as well as the two fricative

16

TOs strategy is adopted by, for example, Moulton (1962: 21), Werner (1974: 40), Wurzel (1981b: 941), Vennemann (1982a: 274), Maddieson (1984), Benware (1986: 57) and Russ (1994:138).

11

pairs /f, ν/17 and /s, TJ, which is the set of segments usually considered for the purposes of a discussion of FOD in the literature (e.g. by Wurzel (1981b: 951) and Wiese (1988: 80)).

One of the aims of this book is to provide a principled account of the distribution of these obstruents with regard to voicing. As already mentioned, I intend to deal both with the dynamic and the static sides of this coin. In other words, I will investígate alternations and distributional restrictions. What exactly these alternations and restrictions are will be discussed in the following sections.

1.2.2.2 Voicing alternations As observed at the very beginning of the present chapter, the application of FOD may give rise to voicing alternations, depending on whether the relevant obstruent occurs in word-final position or not. To see that this cannot be interpreted as an intervocalic voicing process instead, consider the minimal pair Rade - Rate ([taida], 'wheel', dat. sg. [taita], 'council', dat. sg.). The occurrence of intervocalic [t] excludes this possibility. (2) below contains a list of words illustrating the effect of FOD for the five obstruent pairs identified above. A detailed explanation of the arrangement of the data, as well as of the term Fremdwörter and the notational conventions used immediately follows this list. (2)

FOD in word-final position a. Native words Dieb Stab Lob Laub Lied Rad Tod

17

[di:p] [Jtaip] [lo:p] [laup] Dilti [rait] [tort]

[•di:b+a] ['Jta:b+s] [1o:b+a] [laub+a] [liid+u] ['raid+9] ['toid+aj

'thief 'staff 'praise' 'foliage' 'song' 'wheel' 'death'

It may well be the case that what looks as if it should be treated as a labio-dental fricative in German is in fact a glide. There does appear to be some evidence which suggests that /v/ behaves like a hybrid (see e.g. Wiese 1988:93), with properties of both a fricative and a glide. Similar observations have been made regarding Russian M by Jakobson (1978) and Hayes (1984; see also Kiparsky 1985 and Mascaró 1987a for further discussion). Gussmann (1992) also reports that M seems to have a special status in Polish which manifests itself in the fact that it does not always exhibit regular behaviour in the context of devoicing processes. Hall (1992b: 162-168) accounts for the special properties of post-obstruent [v] in German by positing an underlying /u/. This issue, however, is not immediately relevant to the discussion at this point. I shall have more to say about [v/f] in 1.2.2.2.

Kleid Sieg Tag Trog Zweig brav fies Gas Los reis' Haus

[klait] [zi:k] [ta:k] [troik] [tsvaik] [bra:f] [fi:s] [ga:s] [lois] [rais] [haus]

'klaid+e] 'zi:g+a] 'ta:g+3] 'troig+s] 'tsvaig+a] 'bra:v+a] fi:z+3] 'ga:z+9] lo:z+9] 'raiz+a] 'hauz+a]

'dress' 'victory' 'day' 'trough' 'twig' 'good; well-behaved' 'nasty' 'gas' 'lottery ticket' '(I) travel' •house'

ebb' jobb' Kladd' egg' König

[?ερ] [ The rule clearly generates the correct output, but the assumption that underlyingly voiced obstruents are underspecified for voicing on which it crucially depends is somewhat unorthodox.5 As Harris & Lindsey (1995) observe, 'in the normal case, it is the marked term that is lexically present'. So we would expect [ - voice] to be underspecified, not [+ voice], as in the case of Giegerich's account. (Mascaré 1987a, for example, has underspecified [- voice].) Piggott (1988: 161) even goes as far as to claim that 'voiceless obstruents are universally underspecified for the feature value [- voice]'. Be that as it may, (2) represents something of an improvement on an earlier version of the same basic strategy as proposed by Kloeke (1982a: 30), where, as we shall see later, two separate, extrinsically ordered rules have to be used instead. The simplification achieved by Giegerich is a direct consequence of his use of underspecification, which enables him to unify inputs which remained distinct for Kloeke. So far, we have seen that even the apparently so straightforward and familiar rule formulation [- son] -* [ - voice] can represent either a feature-changing (structurechanging) or a feature-specifying (structure-building) rule. Which is applicable depends entirely on concomitant assumptions about the use of underspecifed features. This is, however, not the only aspect of this part of the final obstruent devoicing rule which is worth discussing. The choice of particular binary features ([± tense] instead of [± voice], for example) is another issue which is taken up in some accounts of final devoicing.

2.2.2 The textbook solution in a different guise: [ - son] -*• [+ tense] Although the overwhelming majority of writers on the subject of FOD opt for an analysis which involves the binary feature [± voice], there are some who feel that this is inadequate. I have come across two alternatives involving other binary features. To my 5

This assumption, of course, is not just made by Giegerich {op. cit. ), but also by Walther (1993), for example.

28

knowledge at least, Benware (1986: 66) is the only author who makes reference to [± fortis] when expressing final devoicing, while several others (Kloeke (1982a, b), Giegerich (1989) and Yu (1992b)) prefer to use [± tense] ([± tns] for short) in their rules. It is worth noting, though, that [± fortis] and [± tns] do not merely replace [± voice]. They form part of a pair of features which is taken to be involved in voicing processes and the motivation for which will be discussed below. All authors using [± tns] agree that the other member of the pair is essentially voicing, but neither Giegerich (1989) nor Yu (1992b) refer to it by a feature label, precisely because they consider it to be noncontrastive and predictable, while Benware (1986) includes it as [± voice] in his rules and Kloeke (1982a,b) as [± stiff vocal cords] ([± stVC] for short) in his. Incidentally, Benware's (1986) feature [± fortis] appears to correspond exactly to the other authors' [± tns]. Why is it, though, that these researchers feel the need for two features rather than one? Kloeke (1982a,b), whose influence is acknowledged by both Giegerich (1989: 51) and Yu (1992b: 164f.), was the first to propose this solution, so I shall take a more detailed look at his arguments. The justification for referring to both [± tns] and [± stVC], according to Kloeke (1982a: 30ff.) can be derived from minimal pairs such as those shown in (3). (3)

Tan[tsl?]ar Forschung[s$]aten Sta[tg]asse

Tan[tsp]aar Forschung[st]aten Sta[tk]asse

Theright-handobstruent inside the square brackets in the left-hand column is a voiceless lenis, specified as [- tns, + stVC], while the corresponding segment in the right-hand column is a voiceless fortis, specified as [+ tns, + stVC]. Under Kloeke's analysis, segments which have undergone FOD share the latter specification, i.e. [+ tns, + stVC]. It is, however, not necessary for Kloeke's devoicing rule to specify the values of both features ([+ tns, + stVC]) in its output. He argues that it is really only the feature [± tns] which is distinctive for obstruent pairs. The voicing feature [± stVC] is predictable and, therefore, need not be included in the devoicing rule. Whether a segment is specified as [+ stVC] or [- stVC] depends on two things firstly, whether it is [+ tns] or [- tns] and, secondly, on the environment in which it occurs. Only obstruents which are [- tns] are eligible for [- stVC] status. Lenes (i.e. [- tns] segments), according to Kloeke, are assigned the feature value [- stVC] only when occurring word-initally or between other [- stVC] segments (i.e. vowels and sonorants). So we would expect to find voiced lenes in words such as Bier ([bi:ç], "beer"), Liebe ([li:ba], 'love') and Wende (['venda], 'turn'). It is worth remembering that the voiced lenes ([- tns, - stVC]) are exclusively derived - they do not exist underlyingly. Since only [- tns] segments can be specified as [- stVC], it follows that voiced fortes (*[+ tns, - stVC]) are non-existent. In other words, Kloeke posits a three-way distinction as far as the voicing status of obstruents is concerned. Underlyingly, he differentiates between voiceless lenes ([- tns, + stVC], examples in the left-hand column of (3)) and

29

voiceless fortes ([+ tns, + stVC], examples in the right-hand column of (3)). The third obstruent series, voiced lenes ([- tns, - stVC]), is derived only, as just described. If the distribution of voiced lenes (i.e. segments which are [- tns, - stVC]) is taken care of by a rule which applies after the final devoicing rule, it is no longer necessary to refer to [± stVC] in the context of final devoicing. Simply saying that [- son] -*• [+ tns] is sufficient, as the segment will automatically be interpreted as being [+ stVC], The rules relevant to voicing in German obstruents are reproduced in order of application in (4) and (5). (4)

Final Devoicing

[- son] -»• [+ tns] /

(5)

[- segm]

Voiced Lenis + cons -son - tns

[- stVC] / # _

Kloeke has essentially two arguments to support his claim that three obstruent series are needed. Thefirstis the existence of minimal pairs such as those in (3) and the second the fact that voiceless fortes (i.e. [+ tns, + stVC]), unlike voiceless lenes (i.e. [- tns, + stVC]), are aspirated, so that two types of voiceless obstruents (fortis and lenis) must be distinguished from each other as well as from voiced obstruents (lenis, [- tns, - stVC]). Let me look at each argument in turn. Minimal pairs such as those in (3) justify only the two-way contrast assumed for German by the vast majority of other researchers (see 2.2.1). Distinguishing solely between [+ tns] and [- tns] (or [- voice] and [+ voice], for that matter) would be sufficient. The question is then whether Kloeke's second argument justifies the existence of a third series of obstruents. In other words, is there phonological evidence which suggests that observations about aspiration and vocal cord vibration have to be captured in the way proposed by Kloeke? Are there phonological rules, conditions or constraints which make crucial reference to [± stVC]? If not, then the best course of action to take may be to abandon this feature altogether, which is effectively what both Giegerich (1989) and Yu (1992b: 164f.) do, although they do not explicitly say so. Kloeke (op. cit.) does not seem to have much use for this feature from a strictly phonological point of view either. The only occasions on which he refers to it (apart from the phonetically motivated rule in (5))6 are in a morpheme-structure condition 6

Koeke (1982a: 231) revises this rule later, but the details of this revision are not relevant to the present discussion.

30 (MSC) which captures the voicelessness of obstruent clusters and in a negative syllable structure condition (SSC) which is designed to account for the fact that a syllable boundary never immediately follows a voiced obstruent. The MSC and the SSC are shown in (6a) and (6b) respectively. (6)

a.

b.

If: Then:

[ - son] [ - son] υ υ [+ stVC] [+ stVC]

SSC ~

Coda X

- son .- stVC. The MSC in (6a) is intended to account for the fact that there is no full voicing in lenis (i.e. underlyingly [ - tns, + stVC]) obstruents which immediately follow fortis (i.e. underlyingly [+ tns, + stVC]) obstruents. For example, in a word such as Witwe (['vitya], 'widow'), the coronal plosive is underlyingly fortis ([+ tns]) and the following fricative is underlyingly lenis ([- tns]). According to (6a), both of them also have to be voiceless (i.e. [+ stVC]). There are two points to make in this context. Firstly, MSC (6a) is both redundant in Kloeke's grammar and faced with exceptions. Its redundancy follows from the fact that voiced obstruents (i.e. [ - tns, - stVC]) are derived solely by the Voiced Lenis Rule (5), which can apply to a lenis obstruent adjacent to another obstruent only word-initially. Word-initial obstruent clusters involving voiced obstruents are exceedingly rare and restricted to the sort of exceptions to the MSC cited by Kloeke himself (e.g. [d$]ungel 'jungle', [gd]dansk 'Gdansk', op. cit. : 31). Removing the MSC from the grammar would then have several advantages. The grammar would be less complicated, no exceptions to an MSC would have to be accepted and the correct forms would be predicted by the Voiced Lenis Rule (5) anyway. The second point I want to make about the MSC in (6a) is that its motivation, just as the motivation for the Voiced Lenis Rule (5) is entirely phonetic. No phonological contrast is affected by either (5) or (6a). Thefirstpoint made about the MSC in (6a) also applies to the SSC in (6b). As far as I can see, this condition is redundant, for exactly the reasons cited in the context of the MSC in (6a). This discussion suggests that the inclusion of the feature [± stVC] is probably an unnecessary complication from a phonological point of view and that it may be beneficial to exclude it from the grammar altogether, which would bring Kloeke's approach fully into line with those discussed in 2.2.1. In any case, the essence of his

31

analysis of final devoicing is independent of the presence or absence of [± stVC], because his devoicing rule directly affects only the [± tns] specification of a segment, with decisions about [± stVC] being conditioned by factors unconnected with final devoicing.

2.2.3 Devoicing as delinking/deletion Up to this point we have considered only accounts of FOD which deal with it by means of feature-changing or feature-specifying rewrite rules, without reference to geometric structure. Thinking of FOD, at least to some extent, as loss of structure is, however, a strategy which has become more and more popular in recent years (although it has by no means succeeded in pushing out the more traditional approach altogether). To my knowledge, Mascaró (1987a) was the first to interpret devoicing as a manifestation of reduction (the loss of a phonological property). Specifically, he argued that final devoicing in German involves the deletion of a [+ voice] autosegment from the representation of an obstruent. The complement rule shown in (7) plays a central role in his analysis. (7)

(Obstruent) Complement (Mascaró 1987a: 9) [- son]

[- voice]

The purpose of this rule is to supply [- voice] to those obstruents which have no voicing specification, that is, to underlyingly voiceless obstruents, which are underspecified for voicing, and to those obstruents which have lost their [+ voice] autosegment through deletion. The same strategy is adopted by Yu (1992b: 165ff.) and Hall (1993). Both authors have default rules which correspond very closely to Mascaró's (Obstruent) Complement (7), although Yu follows Kloeke (1982a) in using [- tns] instead of [+ voice], which, as we saw in 2.2.2 can be considered to be equivalent. Yu and Hall differ from Mascaró in interpreting the loss of the underlying [+ voice] (or [- tns]) specification as delinking rather than deletion. Mascaró's reasons for distinguishing delinking (with possible reassociation) from deletion (where no reassociation may occur) relate to languages without final devoicing, so that, given the focus of the present discussion, we can treat Mascaró's approach as fully equivalent to Yu's (1992b) and Hall's (1993). Although some of the ingredients which are combined in the strategy just described may seem quite familiar from earlier analyses (especially the complement rule in (7), which looks indistinguishable from the 1968 feature-changing rule), it does constitute a marked departurefromwhat went before. Its particular strength lies in the motivation of the individual steps. As is widely accepted (including by Chomsky and Halle themselves, see chapter 9 of SPE), the intrinsic arbitrariness of SPE-type rewrite rules is a major problem for phonological theory, because it opens the floodgates to predicting a vast

32

range of phonological structures and processes, only a small minority of which are actually attested.7 In the present context, one could argue, for example, that it would be just as straightforward to write [- son] [+ voice] or even [+ son] -*• [- voice], with nothing in the theory preventing us from expressing these unattested or at least highly marked processes. The very purpose for which marking conventions were introduced was to prevent this sort ofthing. However, marking conventions are really little more than a patching-up device which leaves the fundamental flaw in the theory untouched. The extent to which the same criticism can be levelled at the more recent analysis of FOD proposed by Mascaré (1987a)8 is much more limited. It may be true that the complement rule in (7) is no less of a taxonomic restatement of observed typological facts than Chomsky & Halle's marking convention (1), but the arbitrariness contained in the potentially feature-changing rewrite rule has been reduced through breaking FOD up into several independently motivated processes. Consider each step in turn. Simplifying Mascaró's proposals slightly, we can say that the first stage is the delinking of underlying [+ voice]. This leaves the segment in question without an autosegment representing its voicing status. Using only devices which have been justified in the autosegmental literature (see e.g. Goldsmith 1990 for an overview) for phonological events wholly unconnected with voicing, Mascaró can then account for a broad range of voicing phenomena in a number of languages. Spreading of [a voice] (where a ranges over '+' and '-') from a neighbouring segment then results in voicing assimilation, failure of spreading andfillingin by complement rule results in final devoicing, and reassociation of the delinked autosegment accounts for the behaviour of languages with voicing assimilation but no final devoicing. Which of these events take place in a particular language depends on the settings of two parameters, reduction and spreading. Reduction has three values, viz. deletion (i.e. delinking without the possibility of reassociation), deassociation (i.e. delinking with the possibility of later reassociation) and (i.e. no reduction). Spreading, by contrast, has only the two values '+' and '-'. The number and range of different types of voicing events allowed for by these two parameters is much smaller than that predicted by the rewrite rules discussed earlier, which suggests that Mascaró's proposals are indeed an improvement.

7

8

See Charette 1991 (p. 44), Harris 1990 and Kaye 1989 (pp. 59ff.) for a more detailed discussion of this particular point, and, for example, Goyvaerts 1981a and Foley 1977 for criticism of the fundamentals of the SPE-framework. Incidentally, Wiese (1990) adopts a veiy similar approach, in which the basic steps are more or less identical, although there are differences in the theoretical frameworks on which each analysis is based.

33 REDUCTION

(8)

Deletion

b.

Deletion

c.

Deassociation

d.

Deassociation

e. f.

SPREADING

Voicing effects Word-final devoicing, regressive assimilation: Dutch, Catalan Syllable-rhyme devoicing: German Regressive assimilation: Serbo-Croatian, (Porteño) Spanish No effect: (Nicaraguan) Spanish Limited assimilation: English Same as d.

2.2.4 Constraints on the licensing of a privative [voice] feature The approach to voicing phenomena initiated by Mascaré imposed restrictions on the specification of voicing which can be interpreted as first moves towards privativeness. It, therefore, is not surprising that a fully privative view of voicing in a generative framework was adopted in the context of devoicing fairly soon afterwards. Goldsmith (1989, 1990: 123ff.) assumes that the feature [voice] is privative and argues that this feature can be autosegmentally licensed only under certain circumstances (to be discussed in more detail in 4.3). Suffice it to say for our purposes here that, in his view, a language such as German has codas which are unable to license this feature. As a result, obstruents in syllable codas remain voiceless. A related view is adopted by Lombardi (1991).9 Her central claim is that [voice], alongside two other privative features, [aspiration] and [glottalisation], must be invoked for expressing laryngeal distinctions in consonants and that, in languages with devoicing effects (or 'neutralisation', to use her term), the feature [voice] is only licensed in what is usually described as a syllable onset. This is captured in a positive well-formedness constraint which is in force only in such languages. So, both Goldsmith and Lombardi make reference to a privative feature [voice], whereas Mascaró (1987a) uses the binary feature [± voice] familiar form earlier work. The binary-feature view of voicing allows for the possibility of [ - voice] as well as [+ voice] being active in phonological processes, while the privative approach is designed to exclude the former possibility, thus effectively halving the number of possible phonological events. If, as Lombardi argues explicitly, a privative feature [voice] is sufficient for handling the attested events involving voicing, then replacing

9

See Kenstowicz 1994 (pp. 493ff.) for a useful summary of the central ideas.

34

[± voice] with [voice] is a desirable move, as it simplifies the grammar and helps to further constrain the generative power of phonological theory.

2.2.5 Conclusion In this section I have discussed several approaches to the issue of what devoicing consists in. What has emerged is a progression which moves towards ever greater restrictiveness. This is achieved in two distinct ways. Firstly, the operations employed and, secondly, the primitives affected have become more and more constrained. Early post-SPE analyses involved unrestricted rewriting or feature changing (e.g. Vennemann 1968). These were partly replaced by alternatives making use of one or more independently motivated moves (reduction, spreading; Mascaré 1987a). Even more recently, the notion of the licensing of features is being invoked (Lombardi 1991, Goldsmith 1989, 1990). Both the reduction-and-spreading (Mascaré) and the licensing (Lombardi, Goldsmith) accounts make use of parameterised constraints. As far as the primitives invoked are concerned, there has been a shift from binary features (virtually all analyses) to privative features (Goldsmith 1989, 1990 and Lombardi 1991). Although Mascaré was one of the first to use the term 'reduction' in the context of final devoicing, reduction is only the first step towards devoicing in his account. At the end of the derivation, the reduction effects have been undone again, in the sense that where a segment was unspecified for voice it will have received a new [+ voice] or [- voice] specification by spreading or complement rule. Goldsmith's (1989, 1990) and Lombardi's (1991) strategy is unique among those discussed here in that, although they do not use the term themselves, devoicing does represent genuine reduction in structure for them. A devoiced segment lacks a voicing feature at the end of the derivation, while a voiced segment will have an additional feature, namely the feature [voice].

2.3 The link between the 'what' and the 'where' This brief section provides a link by considering a link. The link it provides is between the two major parts of the present chapter and the link it considers is that between the 'what' and the 'where' of FOD, as captured (or not) in analyses of FOD over the past twenty-five years. Let me begin with the 'textbook' accounts, which use the familiar SPE-type rewrite rule [- son] -»· [- voice] / #. The question I want to focus on is this. What predictions does the statement of the process make about the environment and vice versa? The answer is reasonably clear: none. Everything to the left of the slash is quite independent from the environment stated to therightof the slash in such a rule. By looking at the part of the rule which says [- son] [- voice] we can get no clues as to what to expect in the

35

environment specification (and vice versa, of course). FOD could apply anywhere in a word, initially, medially or finally. To put it differently, we could encounter it in what has traditionally been referred to as a 'strong' position (initially) or in a so-called 'weak' position (finally or medially, e.g. intervocalically,). The reason for this is that the rule [ - son] [ - voice] leaves the question o f what kind of process we are dealing with quite open. If one looks at the change in the redundant features which accompanies this rule and interprets it as [+ voice] -»· [ - voice], then one might venture the guess that FOD is some kind o f strengthening process, since many (but by no means all) consonantal strength hierarchies10 rank voiceless obstruents as being stronger than their voiced congeners. Is this something which is reflected in the environment specification? Or, to put it differently, is there anything in the phonology which tells us that a word-final process is likely to be strengthening? On the contrary - the phonological theory itself makes no claims whatsoever about this. In fact, if one were to look beyond what is built into representations and consider taxonomic statements, one would find that word-final position is a typical 'weak' environment. In other words, segments which occur close to a right-hand word boundary are more likely to lenite than segments occurring in any other environment.11 Some researchers who have investigated final devoicing in these terms seem to be aware of the apparent contradiction this entails (e.g. Lass 1971). Escure (1977), however, may be unique in actually biting the bullet and accepting that either the strength hierarchy or the classification of the final environment as weak must be false. Faced with a wealth of evidence from uncontroversial lenition processes occurring in final position, she decides that it is the strength hierarchy which is at fault. Her solution to the problem is to stipulate that the relevant consonantal strength hierarchy has to be modified in such a way that, in final position only, voiceless obstruents are weaker than their voiced congeners. I shall have more to say on this point in 3.4.31. Further evidence in favour of treating final devoicing as weakening comes from phonetic studies (e.g. Ni Chasaide 1989) which argue that final devoicing, like the uncontroversial weakening processes o f spirantisation and sonorisation, is a result of target undershoot due to certain articulatory and aerodynamic factors a discussion of which would be beyond the scope of this book. In short, when considering the part of the rule intended to capture what final devoicing is ( [ - son] - * [ - voice]), we find practically no information on what type of process is involved and, hence, where FOD should occur. The one prediction one may be able to

10

Consonantal strength hierarchies are the inverse of sonority hierarchies. Those which distinguish between voiced and voiceless obstruents include the ones proposed in Lass 1971,1984 (p. 178), Escure 1977, Hooper 1976 (p. 206), Vennemann 1972,1982a, Lass & Anderson 1975 (p. 159) and Selkirk 1984. In contrasting work, e.g. Kiparsky 1979, van der Hulst 1984 (p. 46), Wiese 1988 (p. 91), Clements 1990, Hall 1992a, 1992b (p. 64) and Roca 1994 (p. 154), the authors eschew this distinction and group all plosives or all obstruents together. Critical comments on the concept of sonority hierarchies can be found, for example, in Harris 1985 (chapter 2), Wiese 1988 (90ff.) and Heike 1992.

11

See Escure 1975, 1977, Harris 1990, 1992 and Harris & Kaye 1990 for evidence.

36

extract from it by invoking strength hierarchies turns out to be false. The fact of the matter is that what we can glean from the left-hand side of the rule contradicts what we find once we take into account the environment specification on the right-hand side (combined with relevant lenition studies). No part of the rule in itself makes any predictions about the nature of FOD. These can only be derived with reference to evidence completely outside the theory itself. This unsatisfactory state of affairs improves somewhat with the advent of approaches to FOD which involve some kind of reduction. Mascaró's (1987a) delinking/deletion account could be seen as implying weakening. This could be a positive result, since, as we have just seen, the most well-known FOD environment, word-final position,12 is a typical weak environment. The assumption that reduction of feature content is to be seen as weakening is actually not something that can be taken for granted, though. Rice (1992), for example, argues the exact opposite. For her, an increase in structural complexity is what accompanies lenition events. As we shall see in 3.3, Government Phonology (see 3.1 for references) is a phonological theory which, in contrast to Rice's framework, represents lenition in terms of loss of structure. If we apply this view to Mascaró's analysis, at least part of what he proposes can be seen as weakening, although we need to bear in mind that the weakening effect is undone again later in the derivation by the application of the complement rule in (7). Goldsmith (1989, 1990) and Lombardi (1991) are the first to persevere with reduction. For them, a privative feature [voice] is absent from voiceless obstruents. Adopting the Government Phonology view of what constitutes weakening, we can then see a connection between the environment and the event. However, we need to bear in mind that no such connection can be established on the basis of what Goldsmith and Lombardi themselves state. We can conclude, for the time being at least, that none of the accounts of FOD under discussion here establish a clear and explicit link between the process and the environment in which it occurs. That setting up such a link would be desirable will become clear when we consider other processes (such as /•-vocalisation and velar spirantisation in German) which apply in exactly the same environment as FOD and try to determine what they and FOD have in common. This will be done in the remaining chapters. For now, let us follow up a remark I made earlier about word-final position being the most well-known FOD environment and investigate where else this process can apply. In other words, we are trying to discover what triggers FOD. This is an issue which has received rather more attention in the literature than the question of what FOD is, so that there are quite a few studies to take into account. I shall group them together according to the kind of analysis they propose and deal with them in roughly chronological order.

12

We shall see later in this chapter and, especially, in chapter 4 that there are other environments in which FOD can apply. All of them can be shown to be weak environments.

37

2.4 Where does FOD occur? 2.4.1 Introduction In this section, I shall be concerned with the question of where FOD occurs in German, i.e. what the triggering environment is. There are essentially four distinct positions which have been adopted in past work and which will be discussed here, three of which are reasonably popular, while the most recent one is represented by only one author. I feel that it is innovative enough to be worth considering, even though it is perhaps too new or too unconventional to have found many followers. The post-SPE account with the longest tradition is that first proposed by Vennemann (1968), which argues that obstruents are subject to FOD when immediately preceding a morpheme boundary. This view has also been adopted by authors such as Wurzel (1970) and Kloeke (1982a, b), although it must be noted that there are important differences both in the spirit and the letter of these three accounts. The implications of the morpheme-final approach, together with possible problems, will be discussed, focusing first on Vennemann 1968 and then on Kloeke 1982a, in 2.4.2. More recent analyses have argued that it is not a morphological boundary which needs to be referred to when discussing the question of where FOD occurs but the syllable boundary. Here, we can distinguish two slightly different strategies. The chronologically older one (first proposed in Vennemann 1972) holds that obstruents devoice when immediately preceding a syllable boundary, while the more recent version of the syllablebased approach (e.g. Vennemann 1978, Rubach 1990) argues that it is segments syllabified into a coda which are subject to final devoicing. Although, at first blush, the two syllable-based analyses seem more or less equivalent, there are some small, but crucial, differences between them, which will emerge in the discussion of both approaches in 2.4.3 (with reference to Wiese 1988 and Hall 1992b) and in 2.4.4 (with reference to Rubach 1990) respectively. In 2.4.5 I briefly consider one of the most recent analyses of the FOD triggering environment, viz. Lombardi 1991. Lombardi argues that German has a positive wellformedness condition, which, roughly speaking, stipulates that [voice] is licensed only in the left branch of a syllable, i.e. in a syllable onset.

2.4.2 FOD applies preceding a #-boundary 2.4.2.1 Vennemann 1968 2.4.2.1.1 The analysis Faced with the question of where exactly FOD occurs, Vennemann (1968: 161-185) sets up a total of seven hypotheses which he confronts with data. The first, that FOD applies word-finally (i.e. ##), is immediately refuted by words such as Liebling (['liipliq],

38

'darling'), where FOD happens word-internally. The second, which identifies the formative boundary (+) as the crucial environment, fares no better 'as is clear from ... staubig ['Jtaub+iç],13 Eises [?aiz+as]' and other examples. This, Vennemann concludes, leaves us with (9)

the option of attributing the cause to (i) (ii) (iii) (iv) (v)

the following voiceless consonant, or the following obstruent, or the following consonant, or the following formative boundary followed by the consonant, or the following #-boundaiy.

Vennemann rules out the first four options one by one. (9i) falls because FOD also applies before voiced consonants, as in sagbar ([ za:kba:ç], 'speakable'), while (9ii) is falsified by counterexamples such as Liebling ([liipliq], 'darling') and Wagnis Q'vaikms], 'venture'), where FOD applies before a sonorant. The third option, by contrast, is too general, as it predicts alternations in paradigms such as ärgern (['feggsn], 'to annoy), ärgre (['îeçgra], 'annoy!') and schneidern (['Jnaiden], 'to tailor'), schneidre (['Jnaidra], 'tailor!'), where the velar and alveolar plosives both remain voiced, even when (apparently) preceding a consonant. The same set of (morphologically complex) items serves to disqualify option (9iv) as well, as witnessed by the underlying representation proposed for schneidre //naid+r+a/. Vennemann concludes that it is a following #-boundaiy which causes an obstruent to devoice. The #-boundaries involved here are inserted by a convention put forward in SPE (Chomsky & Halle 1968: 366), which states that they (10)

are automatically inserted at the beginning and end of every string dominated by a major category, i.e. by one of the lexical categories 'noun', 'verb', 'adjective', or by a category such as 'sentence', 'noun phrase', 'verb phrase', which dominates a lexical category.

A morphologically complex noun, e.g. one derived from an adjective by suffixation, such as Liebling (['liipliq], 'darling'), is assigned the following bracketing and boundaries: [#[#lib#]AliNg#]N. Devoicing then applies after (among other rules) Nasal Assimilation and /g/-Deletion, generating the correct output ##li:p#liq#. However, the proposed devoicing rule in (11) makes the wrong prediction for vowelinitial suffixes, as the data show FOD to be blocked before a vowel. (11)

[- son]

[- voiced] / _ #

Therefore, Vennemann introduces the pre-cyclic #-Deletion Rule quoted in (12). 13

Throughout, I shall adopt the practice of slightly adjusting other authors' transcriptions to make them fit the system used here, in the hope that their substance will not be compromised.

39

(12)

A single internal #-boundaiy is deleted before a vowel.

This rule is supplemented by the convention in (13), which ensures that no single #-boundaries are left behind. (For the purposes of this convention, Vennemann defines the counterpart of a #-boundary as the other member of each pair of #-boundaries specified in SPE-convention (10), which introduces #-boundaries.) (13)

When a #-boundary is deleted, so is its counterpart.

Independent motivation for rule (12) comes from German stress. Vennemann claims that the noun Schneider (['/naide], 'tailor') differsfromthe noun Liebling ([Ιίφΐιη], 'darling') in its stress pattern. The former exhibits [1 -], while the latter has [1 4], As both of them share identical bracketing and #-boundaries ([#[#sneyd#]vir#]N and [#[#lib#]AliNg#]N respectively) and stress is assigned by a rule making reference to #-boundaries,14 this difference in stress behaviour must be due to the deletion of a #-boundaiy from the former noun (Schneider), as stipulated by rule (12), working in tandem with convention (13). Vennemann's SPE-type analysis of FOD can then be reduced to the claim that obstruents are devoiced before a #-boundary, with the proviso that a single internal boundary of this type is deleted before a vowel. Inflected forms and obstruent clusters not involving #-boundaries are not dealt with in the section on devoicing. They are handled by other voicing rules, in particular rule (14) (op. cit. : 73). (14)

Obstruent Cluster Devoicing Rule [- son] [- son] 1 2 12 }

[- voiced]

In this way Vennemann distinguishes between two types of devoicing process. The first is triggered by a #-boundary and has no apparent phonetic motivation, while the second is basically an assimilation process, where two adjacent obstruents are specified as sharing the same value for the feature [± voice], i.e. [- voice]. It is important to note here that Vennemann deliberately restricts the data on which his entire thesis is based to native words. Fremdwörter, he argues (pp. vif., 4ff.) 'require a separate study' (p. vi). It is for this reason that the very general obstruent cluster devoicing rule (14) makes the correct predictions for the majority of the data. There are a number of foreign items (e.g. 14

'Stress Placement Rule: Each string of units enclosed in #, but not containing #, receives exactly one stress, which falls on the first vowel.' (p. 165)

40

Smaragde ([sma'rakcb], 'emeralds') which do not conform to (14), but there are also a few native words (e.g. Mägde ['me:kda], 'maids') which fail to do so. Be that as it may, this analysis enables Vennemann to account for a wide range of otherwise problematic data. Most notably, he can deal with a list of what he describes as 'controversial items'. While speakers of Standard German generally use FOD quite uniformly and consistently (with strong intuitions about its application), there is systematic variation in the pronunciation of these items between two groups of speakers concerning the application of FOD. I will use the term 'variable' rather than 'controversial' for items such as those in (15). For speakers in one group, FOD does not apply (Vennemann calls this group A), while speakers in the second group (Vennemann's group B) consistently devoice where Vennemann's analysis, as presented so far, would predict the absence of FOD. Vennemann's group A corresponds to speakers of Hochlautung and his group Β to speakers of what Vennemann (1972) refers to as 'Standard German, Northern pronunciation' and for which I shall use the term 'Northern Standard German' (NSG).15

(15)

Group A

Group Β

a.

eignen Lügner regnet Segnung

['?aignsn] [1y:gm?] ['reignst] ['ze:gnuq]

[?aiknan] t'lyiknu] ['re:kn9t] ['zeiknuq]

'to suit' 'liar' '(it) rains' "blessing'

b.

ebnen Ebnung

[?e:bnsn] [?e:bnuq]

[?e:pnan] [7e:pnuq]

'to level' 'levelling'

c.

ordnen [?oçdnan] Ordnung ['îoçdnuq]

['?oçtnsn] ['?oçtnuq]

'to arrange' 'order'

(ii)

Adler edler handle Handlung

['?a:dh?] [?e:dle] f'handla] ['handluq]

['faitte] [?e:tlB] ['hantla] ['hantluq]

'eagle' 'nobler' '(I) act' 'act'

(iii)

duslig

['du:zliç]

['duisliç]

'dizzy'

(i)

This difference in pronunciation Vennemann attributes to a difference in syllable division for the two groups of speakers. Where '"unpronounceable" clusters' ([gn bn dn di zl]) would arise from treating both consonants as tautosyllabic, B-speakers insert a boundary

15

Speakers in the latter group often spirantise underlyingly voiced velar plosives. This will be discussed in more detail in 5.4.2.

41

to make them heterosyllabic. Formalising this process is problematic in the orthodox SPE-framework, as it does not recognise the syllable as a phonological unit and, hence, there is no means of referring to syllable-boundaries. In the absence of an actual syllable boundary Vennemann uses the #-boundary instead. This is possible because, he claims, the presence of a #-boundary implies the presence of a syllable boundary (but not vice versa). All he needs to say, then, is that B-speakers (those who treat the two members of the 'unpronounceable' consonant clusters as heterosyllabic and have final devoicing in the words in (15)) insert a #-boundary to separate the two consonants. This boundary insertion is handled by a relatively late rule which only B-speakers have. This rule (op. cit. : 183) is reproduced in (16). (16)

Supplementary it-Introduction Rule [A] #-Boundary is introduced in post-stress position between the members of the 'unpronounceable' clusters /gn bn dn di zl/.

According to this analysis, radle (['raid/tla], '(I) cycle'), for example, is syllabified as by B-speakers, but as ra.die by Α-speakers. The difference in syllabification is due to the fact that B-speakers insert a #-boundary between the plosive and the lateral by applying a rule which is unique to the group of B-speakers (rule (16)). The derivation of radie would then go something like this. Both groups have the underlying representation [[[rad]NVl]vV]v, which becomes [#[#[#rad#]NVl#]vV#]v (by convention (10)) and then [#[[rad]NVl]vV#]v (by rule (12) and convention (13)), and, ultimately, #ra:db#. Bspeakers (but not Α-speakers) now insert a #-boundary to split the 'unpronounceable' cluster. This insertion yields #ra:d#ls#, which, finally, ends up as #ra:t#la# (by (11)). Vennemann seems to have killed two birds with one stone here. Firstly, he has found a way of expressing syllable divisions in terms of the syllable-denying framework of SPE and, secondly, the solution to that problem, the insertion of a #-boundary, also enables him to make the correct predictions for FOD. Whether this approach is really as successful as it seems will be discussed in the next section. radie

2.4.2.1.2 Problems with Vennemann's 1968 analysis 2.4.2.1.2.1 Using a #-boundary to represent a syllable boundary In fact, it is Vennemann's apparently so ingenious way of expressing an apparently syllable-based process in terms of a syllable-denying formalism which causes his analysis to run into trouble. The benefits of using a #-boundary in place of a syllable boundary are bought at considerable cost. First of all, this solution mixes morphological categories (#-boundaries) and phonological categories (syllable boundaries), a strategy which has rather interesting

42

implications. The reason why these implications are not immediately apparent from Vennemann's thesis is that he fails to further examine the predictions made by (16). Recall that SPE-convention (10) introduces #-boundaries to separate major categories from one another. In other words, nouns, verbs, adjectives, noun phrases etc. are each surrounded by a pair of #-boundaries. The prediction in a consistent framework would be that wherever a #-boundary appears there must be something belonging to one of the major categories immediately to its left or its right (unless, of course, there is another #-boundary intervening, which would be the case where two major categories meet). This prediction is not borne out in a number of cases. To see this, consider the predictions made for the words in (15). The boundary insertion rule (16) implies that eig-, lüg-, reg-, seg-, eb-, ord-, ad-, ed-, hand- and dusare the major categories from which the actual forms in (15) are derived by adding the suffixes -neri, -ner, -net, -nung, -1er, -le and -lig, or, more precisely, the suffix sequences -n-en, -n-er, -n-et, -n-ung, -l-er, -l-e and -l-ig respectively. We can investigate whether this claim is correct by looking into the behaviour of some of the suffixes involved here. Taketiienominal suffix -ung and the 1st sg. pres. ind. inflectional suffix -e, for example. Both of them can be attached immediately to the stem (which is identical with the root here), as witnessed by the following forms. (17)

a.

Werb-ung Lad-ung Neig-ung Les-ung

f'veçbuq] [laiduq] ['naiguq] [1e:zuq]

'advertising' 'load' 'inclination' 'reading'

b.

werb-e lad-e neig-e les-e

['veçba] ['larda] ['narga] [le:za]

'(I) advertise' '(I) load' '(I) incline' '(I) read'

If we repeat the same exercise, but with the lexical items in (15), where suitable forms are available, we get the data shown below (transcriptions reflecting NSG). Eign-ung Segn-ung Ebn-ung Ordn-ung Handl-ung

[?aiknuq] ['ze:knuq] f'?e:pnuq] ['?oçtnuq] ['hantluq]

'suitability' Tjlessing' 'levelling' 'order' 'action'

eign-e lüg-e regn-e segn-e

['?aikna] [ly:g9] ['re:kna] f'zeikna]

'(I) suit' '(I) lie* '(I) rain' '(I) bless'

43

ebn-e ordn-e [ver]edl-e handl-e dusl-e

['?e:pna] ['?oçtna] [feç'?e:tla] [hantls] ['duisla]

'(I) level' '(I) arrange' '(I) refine' '(I) act' '(I) get drunk'

The following conclusion can be drawn from the data in (17) and (18). Either the suffixes -ung and -e are attached to the simple stem (i.e. the root) itself, with no other suffixes intervening (as suggested by the data in ( 17)), or they can be attached to a complex stem comprising the root plus a morpheme -/- or -«-. The former interpretation would mean that -/- and -n- in (18) are part of the root.16 Vennemann's bracketing of Radier ([[[rad]NVl]vV]v) suggests that he adopts the opposite position, that is, that -/- and -n- are morphemes which separate the simple stem (i.e. the root) from suffixes such as -ung and -e. If it can be shown thai Vennemann's analysis of Radier, mutatis mutandis, works for the remaining items in (15) as well, then there would be no problem. On the contrary, this sort of bracketing seems to have distinct advantages, since it (a) expresses the relationship between Rad ([ra:t], 'wheel, bicycle') and Radier ('cyclist') and (b) allows the #-boundary insertion rule (16) to capture a valid fact about the morphological structure of these words. The same could be said for Handlung, although the semantic link between Hand ([hant], Tiand') and Handlung ('action') may be just a little more tenuous than between Rad and Radier. However, it seems that it is these two which are exceptional. None of the remaining items in (15) permit the same type of bracketing, as it would create stems for which no independent motivation can be found. There are no inflectional paradigms which can be shown to involve the stems eig-, reg-," seg-, eb-, ord-, ad-, ed-, and dus-. On the contrary, there is positive evidence which points towards final -I and -n being part of the root itself. This evidence comes from the data set in (19). (19)

16

17

eigen Regen Segen eben Orden

[?aigan] ["re:gan] ['zeigan] ['?e:ban] [Toçdan]

'owned by; peculiar to' 'rain' 'blessing' 'level' '(religious) order'

(18b) indicates that the -n- in Lügner is not part of the stem. It seems, then, that whatever the special properties oí Lügner, which make it behave similarly to the remaining variable items, may turn out to be, they are different from those which account for the special status of the other words in (15). I will deal with items like Lügner in more detail in 4.4.2.3. This claim may appear to be incorrect, as reg- is, in fact, attested, cf. Reg-ung (['re:guq], 'movement) and reg-e ([rc:ga], '(I) move'). This, however, is different from the stem associated with the meaning 'rain', and, therefore, no counterexample.

44

edel Handel Dusel

['?e:dal] 'noble' [handal] 'trade' [duizal] 'dizziness'

There is not a single form in the entire inflectional paradigms of the lexical items in (19) which lacks -/ or -n. If, however, the lateral or the nasal did constitute a morpheme which is separate from the stem, then one would expect precisely such forms to exist.18 The argument just put forward is based on the assumption that the #-boundaries inserted by the supplementary #-boundary insertion rule (16) have the same status and function as those inserted by convention (10), namely to mark out lexical categories. This, however, is not the case. Vennemann's rule (16) actually makes no claims about morphological structure at all. It may be inserting the same type of boundaries as convention (10), but their status is completely different in the context of rule (16). Here, they are not meant to have any implications for the morphology. Rule (16) is simply a readjustment rule, so that the morphological argument developed against it in the preceding paragraphs loses much of its force. However, the mere fact that Vennemann has to invoke readjustment rules such as (16) weakens his analysis, since, as Mohanan (1986: 132) observes, in a theory which makes use of ad hoc readjustment rules (such as (16) in this case), 'an attempt at principled boundary assignment as in [(10)] becomes entirely meaningless'. Boundaries quite generally become little more than diacritics. Even if we accept that the supplementary #-boundary insertion rule (16) is simply a readjustment rule which makes no claims about morphological structure (which, as I have shown, would be incorrect anyway), it still has the considerable disadvantage of being an ad hoc device which, to make matters worse, deprives the original #-boundary insertion convention (10) of its meaning. Having discussed the specific problems arising from the variable items, I will now turn to a much larger data set, that of vowel-initial suffixes.

2.4.2.1.2.2 Deleting a single internal #-boundary before a vowel Consider the two adjectives kindlich (['kintliç], 'child-like') and kindisch (['kindij], 'childish'). Most phonologists (including Vennemann)19 would syllabify these as kindlich and kindisch by the Maximal Syllable Onset Principle (see, for example, Selkirk 1982). According to this syllabification, FOD should apply in kindlich, but not in kindisch,

18

As already observed in footnote 16, Lügner is an example of a case where such a morpheme is present, as witnessed by Lüg-e ('lie') and lüg-en ('to lie'). In the nominative singular of the noun and in the infinitive form derived from the same stem, the -n- present in the agent noun Lügner is missing. By contrast, none of the words in (19) have such -«-less or -/-less forms, which can only mean that -/ or -n are part of the stem. This, I believe, can be inferred from several of the observations he makes in Vennemann 1982a

45

provided, of course, that it is treated as a syllable-final process, which, after all, is the idea of Vexinemann's 1968 account. This is exactly what we find. Given, however, that the SPE-formalism prevents Vennemann from directly referring to syllable-boundaries and forces him to use #-boundaries instead, he has to get rid of the #-boundaiy, introduced by convention (10), in kindisch (but not in kindlich, where FOD applies) in some way. The relevant derivations would go roughly like this. The underlying representations20 [[kind^lich]^ and [[kind^isch]^ are converted to [#[#kind#]Nlich#]Adj and [#[#kind#]Nisch#]Adj respectively by convention (10). The devoicing rule (11), which applies before a #-boundary can apply to kindlich, but its application to kindisch would generate the wrong output. It is for this reason that Vennemann needs a rule which deletes an internal #-boundary before a vowel. This is done by rule (12), repeated here as (20) for convenience. (20) A single internal #-boundary is deleted before a vowel. It is undeniable that this rule makes it possible for the correct output to be generated. Its most obvious shortcoming is that it is not particularly well motivated. For one thing, it lacks phonetic plausibility (see e.g. Hyman 1975: 97f.). Phonetic implausibility by itself, however, is not necessarily a damning indictment for any rule, as witnessed by the abundance of phonetically implausible rules in SPE itself. It only really becomes a problem if there is no independent motivation21 for the rule, that is, if the sole motivation ofthat rule appears to be that it can make a particular analysis work. Such independent motivation, Vennemann claims, comes from German stress facts. He presents a set of stress assignment, reduction and adjustment rules to support his analysis of FOD (including the #-boundary deletion rule in (20)). In my view, this set of stress rules makes wrong or dubious predictions - even for some of the words quoted by Vennemann himself. Here are but two examples to illustrate my point. Firstly, his analysis requires the deverbal suffixed noun Labsal (['la:pza:l], 'refreshment') to be derived identically to the true (nonce) compound noun Labsaal (['la:pza:l], 'refreshment hall') for the stress patterns to be correct, i.e. identical. The difference in morphological structure makes this impossible, as the stress rules refer to morphological bracketing. To see what happens consider the underlying representations involved, viz. [[lab]vsal]N (Labsal, i.e. lab- 'to refresh' plus the nominal suffix -sal) and [[lab]v[sal]N]N. These become [#[#lab#]vsal#]N (Labsal) and [#[#lab#]v[#sal#]N#]N (Labsaal) respectively by convention (10). Vennemann's stress rules would generate the stress pattern [1 4] for Labsal, but [1 3] for Labsaal. Vennemann considers these non-

20

21

I have used orthographic rather than phonological representations here, as I am not entirely certain what Vennemann's chosen representations would be in these cases, and I would like to avoid accidentally misrepresenting his views. See e.g. Kenstowicz & Kisseberth 1979 (pp. 150-153) for discussion.

46

identical stress patterns as incorrect, though, which is why he attempts to rectify the situation with the help of the #-introduction rule shown in (21). (21 )

it-Introduction Rule A ¿/-boundary is introduced before the suffixes -sal, -sam,-bar, -los,...

This rule generates a stress pattern for Labsal which is identical to that of Labsaal, but this appears to be its only raison d'être, as Vennemann is unable to provide independent motivation for it. To sum up my first point of criticism regarding the stress rules intended to underpin the rules required for FOD,flieserules make incorrect predictions and have to be patched up with rescue-rules, such as (21), for which no independent motivation can be found. My second point is that, in other cases, the stress rules make false predictions which force Vennemann to treat certain suffixes as exceptional. The deverbal noun Labung ([laibuq], 'refreshment' - derived from lab-, but this time with the suffix -ung) is a case in point. It is assigned the stress pattern [1 -], which would entail a reduction of the short unstressed vowel to schwa (which is motivated elsewhere in Vennemann's thesis). This, however, does not happen, and Vennemann is forced to treat the suffix -ung (along with several others) as exceptional with regard to vowel reduction. To sum up, Vennemann's attempt to prop up his #-boundary deletion rule (12) (reproduced as (20)) by reference to stress fails because his stress rules are incorrect predictions for stress assignment.

2.4.2.1.2.3 Dealing with variable items: 'unpronounceable' clusters My third andfinalpoint of criticism of Vennemann's SPE-type analysis of FOD concerns his way of dealing with the so-called 'controversial items' (listed in (15)). According to his analysis, speakers of group Β (those who have FOD in these items) find it necessary to insert a #-boundaiy to break up an 'unpronounceable' cluster, while speakers of group A (those who do not devoice in these cases) apparently have no difficulty in treating these 'unpronounceable' clusters as tautosyllabic. The question arises why this difference between the two groups should exist. Vennemann's choice of the term 'unpronounceable' in this context suggests that we may be dealing with a physical or physiological factor which makes it impossible for B-speakers to articulate these clusters tautosyllabically. On closer inspection, however, this view becomes untenable. If it really were the case that Α-speakers are in some way better vocal gymnasts, the fact that entire speech communities must suffer from some kind of lingual affliction would have to be explained. This is clearly not what Vennemann intended. I believe that he is referring to a type of phonotactic, i.e. phonologically motivated, constraint. Let us consider the implications of the claim that Α-speakers treat these 'unpronounceable' clusters as tautosyllabic in greater detail. If Α-speakers do indeed syllabify, say, radie as ra. die, the most obvious conclusion to be drawn from this would

47

be that dl as well as gn bn dn zi are licit syllable onsets in their dialect. Although Vennemann posits underlying representations where the clusters are broken up by a vowel (e.g. [[ordVn]vuNg]N for Ordnung and [[[rad]NVI]v]V]v for radié), the fact remains that the plosive and the sonorant become adjacent through the application of the syncopation rule deleting the vowel which separates them in their underlying representations. This, effectively, makes them syllable onsets, although this notion obviously does not exist in SPE. Vennemann himself interprets them in this way by calling them syllable-initial clusters in Vennemann 1972, where he is in a position to refer to syllable boundaries. There is one problem with this analysis. If these clusters were indeed acceptable syllable onsets for Α-speakers, one would expect them to behave like other onsets, that is, to occur word-initially as well as word-internally. In Vennemann's variable items in (15) (and others like them) they only occur word-internally and, in fact, it appears that, with few exceptions, they are never found word-initially.22 Possible counterexamples to this statement, however, can be found in Bavarian-Austrian dialects, which have forms such as [dlait] (die Leute, 'the people") and [dnorxt] (die Nacht, 'the night'). Do dl and dn occur word-initially after all, at least in these dialects, and, perhaps, also in Hochlautung? The answer to this question could only be yes if it were impossible for a vowel to surface, say, in careful pronunciation, between the stop and the following lateral or nasal. This, however, is not the case. It seems that there is a vowel underlyingly present, since, in amore formal style, the pronunciations [di'lait] (die Leute, 'the people') and [di'no:xt] (die Nacht, 'the night') are also attested. By contrast, no vowel can ever be heard breaking up the offending clusters in words such as Adler or Ordnung (or others like them). Quite irrespective of any theoretical interpretation one could apply to the Bavarian-Austrian forms and Vennemann's variable items, it is clear that the two are not comparable and that in the former case there is a vowel (underlyingly) present, while in the latter there is not, at least not one which can ever be made to surface in the same word, even when a different speaking style is adopted. In other words, it appears that Vennemann's 'unpronounceable' clusters do not occur word-initially in German. Some authors, such as Wiese (1988: 93f.) and Hall (1992b: 71), actually go as far as to set up syllable structure conditions which exclude at least some of these clusters from any syllable onset, both word-initially and medially. The defective distribution of these clusters is not recognised as being particularly important in Vennemann 1968 (in fact, it is only noted in passing), but a different view is taken by Vennemann in his 1972 paper on the theory of Syllabic Phonology (1972). He considers that this defective distribution is in conflict with a universal principle, which he calls the Law of Initials and phrases as follows. (22)

22

Law of Initials (Vennemann 1972: 11 ) Medial syllable-initial clusters should be possible word-initial clusters.

There are some words, though, which begin with [gn] (about 25 in all). These will be discussed in detail in 4.4.2.4.

48

In Vennemann 1988, the Law of Initials is further refined to allow for apparent exceptions. (23)

Law of Initials (revised version) (Vennemann 1988: 32) Word-medial syllable heads [i.e. syllable onsets/WGB] are the more preferred, the less they differ from possible word-initial syllable heads of the language system.

Both versions are clearly designed to express a universally valid insight23 and both are in conflict with Α-speakers' pronunciations of the variable items. Such a conflict leaves two options. Firstly, one could, as Vennemann does, simply state that the variable items (and others like them) are exceptions to this law. Secondly, one could try to find an alternative explanation for the existence of these apparent clusters which does not involve treating them as onsets. Let me discuss the two options in turn. The first, interpreting the unpronounceable clusters as syllable onsets and hence as exceptions to the Law of Initials, has some undesirable implications. One of these is that a general (even universal) constraint on onsets has to be replaced by one or more very specific ones for Α-speakers, stipulating that certain clusters are permitted wordintemally (under certain circumstances only, as we shall see below), while they are ruled out word-initially. One aspect of this is that (as Rubach (1990: 83) points out) devoicing in foreign borrowings or proper names can only be handled with a great deal of individual marking and other machinery, such as #-boundary introduction rules for Α-speakers as well as B-speakers (but in different environments). The sort of problems which are likely to arise from this have already been discussed in 2.4.2.2.1. If, however, the second option were adopted and the Law of Initials applied without exception, the devoicing facts relevant to Fremdwörter (as discussed in Vennemann 1978) would fall out as a result of this law. This approach would, of course, make it necessary for the 'unpronounceable' clusters which are allegedly onsets for A-speakers to be analysed as something other than onsets. This is not possible in the orthodox SPEframework (but it is in others, as we shall see in 2.4.4 and in chapters 3 and 4), which is why Vennemann is prevented from choosing this, otherwise preferable, option. The case of the 'unpronounceable' clusters is a good example of how a useful insight (the Law of Initials) can be derailed by an inadequate theoretical framework.

23

To cite just one piece of evidence in support of the Law of Initials from a non-Germanic language, Mascaró (1986:173) observes that dl cannot be an onset in Catalan for two reasons. Firstly, just as in German, the cluster cannot occur word-initially and, secondly, the d can assimilate to the / (i.e. dl It). The assimilation process involved here, however, normally only permits syllable-^na/ segments to assimilate to syllable-initial ones. In other words, if dl were indeed a tautosyllabic cluster, as implied by the analysis of dl as an onset, assimilation should be blocked.

49

2.4.2.1.2.4 Summary To sum up, Vennemann's 1968 analysis of FOD, which involves the claim that obstruents devoice before a #-boundary, is problematic for three reasons. Firstly, the use of a #-boundary instead of a syllable boundary, which is forced on him by the syllabledenying framework of SPE, leads to an infelicitous intermixing of morphological and phonological categories. This flaw, however, does not actually become apparent until one examines the variable items in (15) in more detail. The insertion of an additional #-boundary by rule (16) for B-speakers, which is specially introduced to deal with these items, has the disadvantage of being an ad hoc device which renders convention (10), which inserts these boundaries to mark lexical categories, completely meaningless. Secondly, the use of a #-boundary instead of a syllable boundary causes problems with making the correct predictions for the application of FOD itself, since the two types of boundary do not always appear in identical places, even in native German words. Vowelinitial suffixes (which would be completely straightforward for a truly syllable-based analysis) necessitate a phonetically implausible rule which deletes internal #-boundaries in order to block the application of FOD before a vowel. Vennemann's attempt to provide independent motivation for this #-boundary deletion rule is unsuccessful because the stress rules adduced for this purpose are dubious themselves. Finally, Vennemann's treatment of the variable items exposes further weaknesses in his analysis. It crucially hinges on the claim that certain groups of speakers (A-speakers) have syllable-initial clusters of the type dl, dn, bn etc. This claim is very difficult to support, as these clusters do not occur word-initially (as do other syllable onsets). In fact, they appear to be restricted (in apparently syllable-initial position) to the very type of words for which they were posited.

2.4.2.2 Kloeke 1982 (and Wurzel 1970) 2.4.2.2.1 The analysis On the face of it, Kloeke's (1982a: 128ff., 1982b) account is very similar to Vennemann's 1968 approach, as both make crucial reference to morpheme boundaries. However, in spirit the two are rather different. Vennemann's underlying insight, as far as one can tell, was that FOD applies syllable-finally, and he used the #-boundary only because reference to the syllable was impossible in the orthodox SPE-framework in which his thesis was couched. Kloeke (1982a: 130f.), by contrast, explicitly rejects the possibility of connecting FOD with syllabification by pointing to the list of word pairs in (24). He argues that a syllable-based approach would be unable to capture the distinction between the items in (24a) and those in (24b), all of which contain underlyingly voiced bilabial and velar plosives. It is for this reason that he embraces an approach which treats morpheme structure as the crucial factor.

50 (24)

a.

nefbliç] giefbliç] weiBschnâ[bliç] hû[gliç] zweiflii[gliç]

'foggy' 'many-gabled' 'white-beaked' "hilly' 'two-winged'

b.

erhe[pliç] lie[pliç] buchstâ[pliç] fu[kliç] klû[kliç]

'considerable' 'lovely' 'literally' 'rightly' 'wisely'

The FOD rule (Kloeke 1982a: 132, 207, 1982b: 172) which captures this was first reproduced as (4) in 2.2.2 and is repeated here as (25) for convenience. (25)

Final Devoicing

{

[ - segm] a. [+ cons] > b. [ - son] c.

Rule (25) says thai obstruents devoice24 before a boundary (of whatever sort),25 provided the boundary is followed by another boundary, a consonant or a laryngeal glide. Thus Kloeke's 1982 analysis is virtually identical with Wurzel's 1970 account, except for the fact that the relevant part of Wurzel's rule reads as shown in (26).

^

[+ obstr] -> [ - voice] /

It says that obstruents devoice before a #-boundary if that boundary precedes another #-boundary or anon-syllabic segment. There are two main differences between the two rules. The first is that the third part of Kloeke's rule makes specific reference to devoicing before h and glottal stop (both of which he interprets as [ - cons, - son]) in (25c). However, as he himself observes (1982b: 172), this part may be redundant, as suffixes with an initial A (e.g. -heil and -haft) or glottal stop (e.g. -artig,26 which is probably not a suffix at all, but part of an adjectival compound) seem to be more 'independent' than other suffixes. This fact could be expressed by having two boundaries separating them from the root. In that case, (25a) would apply. The second important difference between (25) and (26) is that Wurzel, unlike Kloeke does not assume that FOD is involved in devoicing effects which can be observed in

24

25 26

Strictly speaking, it says that obstruents become tense (rather than voiceless), but, as shown in 2.2.2, this interpretation cannot be justified. In (25a), the sequence [ - segm] [ - segm] represents a word boundary (#). The semi-word -artig, as Kloeke (1982b: 172) describes it, may appear to be vowel-initial, but it is in fact pronounced with an initial glottal stop. The role and behaviour of glottal stop will be discussed in 4.3.

51

inflected forms. For him, as for Kloeke, inflected forms such as Hunds ([hunts], 'dog', gen. sg.) or gibt ([gi:pt], '(he/she/it) gives') contain only a +-boundary, so that the structural description of (26) is not actually met by these forms. (In Kloeke's case, by contrast, rule (25) can apply here, as it is less specific about the boundary type.) Wurzel assumes that genuine FOD is not involved here. Instead, he argues in favour of a regressive voicing assimilation process which is captured in another part of the same rule, which is shown in (27). (27)

[+ obstr] -> [- voice] /

[- voice]

So, for Kloeke any kind of morpheme boundary can trigger FOD, while for Wurzel only a phonological word boundary (i.e. ##) and a simple word boundary (#) can have this effect. It is for this reason that Wurzel needs to employ a regressive voicing assimilation rule as well as devoicing rule which applies before a (#)#-boundary. To sum up, roughly speaking, both Wurzel and Kloeke argue that obstruents are devoiced before a morpheme boundary, except where that boundary is followed by a vowel. In what follows, I shall focus exclusively on Kloeke's account, as it is much more detailed than Wurzel's. This is not to say that there are no problems with Wurzel's analysis (see footnote 27). For example, one may wonder whether it is appropriate to fragment events connected with voicing as heavily as Wurzel does in his rule, which, in its full form, contains no fewer than four sub-parts.

2.4.2.2.2 Problems with Kloeke's analysis Essentially, the problems with this approach are very similar to those discussed in 2.4.2.1.2, with reference to Vennemann's 1968 analysis, which, although intended to capture the syllable-final nature of FOD, nevertheless expressed the process in terms of #-boundaries. Kloeke is well aware of some of these problems.27 He discusses two types of possible counterexamples to his analysis, Fremdwörter and words which would fit into Vennemann's list of variable items. Let me begin with the Fremdwörter. Kloeke (1982a: 133) argues that most of them simply follow the generalisation that obstruent clusters in German are voiceless (i.e. [+ stVC]). Whether it is really appropriate to invoke this generalisation (as captured in the MSC (6) discussed in 2.2.2) in order to account for forms such as Simba[py]e ('Zimbabwe'), Rufkfrjy ('Rugby') or Wi[ky]am ('wigwam') as opposed to Simbafpvje ('Zimbabwe'), Rufkbjy ('Rugby') or Wi[kv]am ('wigwam') is not immediately obvious. After all, as argued in 2.2.2, reference to [± stVC] is redundant from a phonological point

27

Wurzel (1970) has nothing to say about any problems with his analysis, but see Yu 1992b (pp. 173ff.) for critical comments which, in my view, are only partly justified.

52

of view, as no contrasts are affected. This, incidentally is also illustrated by the foreign borrowings just mentioned. Be that as it may, the real issue is how Kloeke proposes to account for the fact that the plosives in these forms are not only voiceless (i.e. [+ stVC]) but also fortis (i.e. [+ tns]). Taking Rugby, the name of a British town and a ball-game, as an example, we notice that the plosive is voiced in the source language (['jAgbi]), but not in German. The same is true of other non-native items which are structurally similar.28 Kloeke (1982a: 133) claims that, by dint of Vennemann's (1972) Law of Initials (22), the pronunciation of these words in German is expected to contain obstruents which are specified as [u tns], that is [+ tns] (p. 87). I can see no reason for relating this statement to the Law of Initials, as [kb] is no more a permissible word-initial onset than [gb]. A possible solution might be to posit morpheme boundaries where German can provide no morphological evidence for them (e.g. Simbab+we). However, this hardly seems worth entertaining seriously. As Kloeke provides no other arguments in this context, it appears that he is unable to account for the fact that Fremdwörter exhibit this kind of devoicing effect. The fact that Kloeke finds it necessary to posit an underlyingly fortis ([+ tns]) coronal plosive in a word such as Adverb (p. 51) suggests that my interpretation is correct. He is not in a position to offer an analysis which captures this significant generalisation. A further challenge to his analysis of FOD comes from words which are structurally similar to Vennemann's variable items (see (15) above). He (1982b: 172f.) proposes to deal with all of those which involve an obstruent preceding / or η as follows. In words such as Handlung (['handluq], 'action') or Bildner (['bildne], 'sculptor'), both of which are characterised by the absence of FOD for Α-speakers,29 the morphological structure is Hand-l-ung and Bild-n-er respectively, that is, they involve a morpheme -/- or -«-. The presence of such a morpheme is what distinguishes these particular forms from others which are based on the same root, but where FOD applies in all dialects (in Vennemann's terms, for both A- and B-speakers). The adjective hand-lich (['hantliç], Tiandy') and the noun Bild-nis ([biltnis], 'portrait') are examples of such words. The fact that FOD does not apply to words with the morphemes -I- and -n- (such as Handlung and Bildner) is expressed by the readjustment rule reproduced in (28). (28) [u R. (25)] - [- R. (25)] / _ [- segm]

+ cons + son [- segm]

The unmarked value for FOD is [+ R. (25)], but where the SD of rule (28) is met (i.e. where a word contains the morphemes -I- or -«-), rule (25) will be blocked. This generates the correct output for Α-speakers. For this solution to make the right

28 29

See the relevant list in 1.2.2.3 for further examples. In the case of Bildner (and other similar words involving -«-), I have found considerable variation in my own work with Α-speakers. The same speaker pronounced this word [bildne] and [biltne] in a single session. See 4.4.2.3 for further discussion.

53

predictions for B-speakers, all that has to be said is that B-speakers do not have readjustment rule (28). This readjustment rule undeniably captures the facts, but it is hard to see how any independent motivation for its existence could be found. In any case, it seems that Kloeke has overestimated the number of words this rule would apply to. There are, for example, no good reasons for treating a word like Handlung as Hand-l-ung rather than Handl-ung?0 Conclusive evidence in favour of Handl-ung was presented in 2.4.2.1.2. This is actually good news for Kloeke because it means that the readjustment rule (28) has to do a lot less work. If the root is handl-, then the SD of the devoicing rule (25) is not met anyway and Α-speakers' pronunciations are what we would expect. Words like Bildner, of course, would still have to be accounted for by the readjustment rule. B-speakers' pronunciations, on the other hand, will prove problematic. Simply removing die readjustment rule from their rule inventories would no longer do the trick. Kloeke would have to show that for these speakers there is indeed a second morpheme boundary in words like Handlung and Ordnung, which, given the evidence in 2.4.2.1.2, will probably turn out to be impossible. The same problem arises in the context of proper names and a different set of Fremdwörter (to be dealt with in 4.4.2.4), such as Egmont, Bodmer, Magma and Dogma. In Hochlautung, all of these contain voiced obstruents, which is what Kloeke's account would predict. The voiceless realisations of the same obstruents which distinguish Bspeakers from Α-speakers, however, cannot be dealt with in a straightforward manner. There is no evidence whatsoever for internal morpheme boundaries, so that the analysis just considered (and rejected) in the context of Handlung and Ordnung would be ruled out rightfromthe start. Kloeke himself is aware of these pronunciations involving FOD and suggests that, due to the fact that clusters such as [gm] and [dm], as they occur for Α-speakers in Egmont, Bodmer, Magma and Dogma, are in conflict with Vennemann's Law of Initials (22), 'Demarkierung [demarcation]' (p. 133) is to be expected. That is, a voiced obstruent becomes voiceless, following the 'generellen Tendenz, Konsonanten im Cluster zu demarkieren [general tendency towards the demarcation of consonants in a cluster]' (p. 131). What the phonological mechanism involved in this demarcation may be remains entirely open. Kloeke justifies omitting any further consideration of the matter by stating that there is a fundamental difference between such cases and FOD proper, which lies in the fact that no alternation is involved in the former. This is true of proper names and certain borrowings but it certainly is not of other items on the list Kloeke presents in this context, viz. ebnen and ordnen. Here, there is indeed a voicing alternation for B-speakers, as illustrated in 2.4.2.1.2.1, where ebnen (['?e:pnan], 'to level') and ordnen ([ îoptnsn], 'to arrange') were contrasted with eben ([?'e:ban], 'level') and Orden (['toçdan], '(religious) order'). To sum up, any analysis which treats FOD as a morpheme-final process necessarily shares most of the disadvantages of Vennemann's 1968 approach, as already discussed in 2.4.2.1.2. Unlike Vennemann, Kloeke bases his analysis on Fremdwörter as well as 30

The form Handl-ung is proposed by Rubach (in (47)), for example.

54

native forms. However, it turns out that his account does not capture the generalisation that the left-most plosive in words such as Rugby must be fortis ([+ tns]). His appeal to Vennemann's Law of Initials in this context appears to have no relevance to the data at hand. Vennemann's variable items pose another problem for Kloeke's analysis, as some of the stems predicted by Kloeke do not seem to exist. Also, the proposed readjustment rule (28), which handles these words, cannot be motivated independently and one would have to show that the variable items differ in their morphological structures for Α-speakers and B-speakers. Finally, Kloeke is not in a position to offer a phonologically founded account of the mechanism which enables B-speakers to apply FOD to items such as Egmont, Bodmer, Magma and Dogma.

2.4.3 FOD applies immediately before a syllable boundary 2.4.3.1 Introduction The accounts to be discussed in this section as well as in 2.4.4 at least liberate the syllable from its Cinderella existence and allow it to marry the Prince, that is, to play a crucial role in the analysis of FOD (as well as a number of other phonological events, see e.g. Hall 1992b). What distinguishes the accounts to be discussed here from those which will feature in 2.4.4 is the fact that the former focus on the syllable boundary, while the latter consider the syllable coda to be the crucial environment. Some authors, especially of early studies or introductory textbooks, propose both, so to speak, without clearly distinguishing between them, usually by not actually formalising the process (e.g. Moulton (1962:45) and Benware (1986: 66)). Others put forward ideas which, although making crucial reference to the syllable, are rather difficult to fit into one of these two categories. Giegerich's (1989, 1992) strategy involves segments which are syllabified at the right edge of the syllable, but he recognises neither explicit syllable boundaries nor a coda constituent. I shall make no further reference to accounts of this kind, as the most important issues arising in this context can be discussed more straightforwardly on the basis of some of the remaining publications.

2.4.3.2 Wiese 1988 and Hall 1992b: the analysis There are a number of analyses available which argue that the appropriate environment for FOD is immediately to the left of a syllable boundary (including Vennemann 1972, 1978,1982a, Wurzel 1981a, b, Wiese 1988, 1991, Ramers 1992 and Hall 1992b, 1993). For the purposes of the present discussion I shall, however, only focus on Wiese 1988 and Hall 1992b, as these seem to me to be the most detailed and most sophisticated analyses listed here. Perhaps more importantly, theoretical issues arising from more or less all the other approaches are also raised by Hall 1992b and Wiese 1988.

55

The actual rule formulation is virtually identical for the two authors, with the minor exception that Wiese (1988: 80) uses a feature [± obstruent] where Hall (1992b: 53) has [± sonorant], Hall's rule then reads as in (29). (29)

[ - son]

[ - voice] / _ ]„

The differences between Hall and Wiese lie in concomitant assumptions about syllabification and about the domain of application of (29). For Hall, the rule is postcyclic, that is, it applies as part of the lexical rule component, but not cyclically,31 whereas Wiese interprets the rule as postlexical. The standard Lexical Phonology32 assumptions about what this implies are the following. For Hall, the rule may be subject to lexical exceptions and it must respect Structure Preservation (introduced in Kiparsky 1982b), while Wiese's rule is expected to apply across the board and in a non-structure preserving fashion. I shall begin with a brief overview of the kind of machinery involved in deriving the kind of syllable structure which will enable (29) to apply under the right circumstances and illustrate how it works. This will be followed by a discussion of potential problems or controversial issues. Wiese (1988: 60) assumes a flat syllable structure (following Clements & Keyser 1983) which produces the syllable template for German shown in (30). (30)

σ

C C V c

c

This is combined with the Nucleus Condition (31) which ensures that all German vowels are underlyingly long or followed by an ambisyllabic consonant (p. 67). (31)

Nucleus

Λ

V

c

The bare bones of (30) and (31) are fleshed out by positive and negative Syllable Structure Conditions (pp. 89ff.), which, among other things, rule out syllable onsets such as */tl/ and */dV. The sonority hierarchy shown in (32) which Wiese proposes for German also has an important role to play in this context.

31

32

See Booij & Rubach 1987 for a detailed discussion of the differences between postcyclic and postlexical rules. See e.g. Kiparsky 1982a, b, 1985, Kaisse & Shaw 1985 and Mohanan 1986.

56

(32)

ι 1 1 1—I 1 1 plosives fricatives nasals l\l Irl high vowels vowels

Syllabification, which takes place throughout the lexicon and is subject to change in the course of the derivation (p. 85), takes these SSCs and the sonority hierarchy into account to the extent that the syllables generated must conform to the Sonority Sequencing Generalisation (33) and may not normally violate an SSC. (33)

Sonority Sequencing Generalisation (Selkirk 1984: 116) In any syllable, there is a segment constituting a sonority peak that is preceded and/or followed by a sequence of segments with progressively decreasing sonority values.

With reference to the above, the syllabification principles (34) can then generate the majority of well-formed German syllables (quoted, in my translation, from p. 86). (34)

a. Associate V with a local sonority maximum b. Associate from V to the left, as far as possible. c. Associate from V to the right, as far as possible.

In order to accommodate the attested German syllables not catered for by the devices listed so far, Wiese also allows for the existence of extrasyllabic segments (pp. 94ff.), i.e. syllable-initial and syllable-final appendices. Specifically, he argues that a single position may occur extrasyllabically at the beginning of a word and that this position is filled by /// before /p/ and IM and by /s/ before Ikl. Word-finally, l\l, /s/ and /st/ are to be treated as appendices. The basic strategy adopted by Hall (1992b) is very much the same, although there are some differences in the details. For example, Hall's negative SSCs (p. 71) are much more elaborate, and he assumes that ambisyllabicity 'is a phenomenon that plays no role in the lexical as well as most of the postlexical phonology' (p. 49). Furthermore, Hall's (1992b: 64) sonority hierarchy has fewer divisions than Wiese's, a move which follows proposals by Clements (1990). Hall's sonority hierarchy for German is shown (without the feature matrixes identifying each segment type) in (35). (35)

1

2

obstruents nasals

3 1

4

4.5

5

r

vowels

Similarly to Wiese, Hall allows for a type of stray segment adjunction, including coronal fricative adjunction (pp. 74ff.). Unlike Wiese, he employs a Syllable Structure Algorithm (36) which, when applied as a single block in the order given throughout the lexical and postlexical phonology, generates the majority of well-formed syllables of German. Again, like Wiese, Hall considers «syllabification a necessity, although it may be worth noting that only the CV Rule may alter existing syllable structure (pp. 59ff.).

57

(36)

Syllable Structure Algorithm (SSA) a. Nucleus Placement

Ν Χ

b. Syllable Placement a

c. CVRule a Ο ι Χ

Ν (X)

[+ cons]

[- cons] c. Onset Rule

O . - Ί Χ χ

Ν

d. Coda Placement a Ν

C

e. Coda Rule a r \ Ν C l·^ XI X X

The SSA (36) alone, however, is insufficient to assign the correct syllabification in all cases. For this reason, Hall (1992b: 86) introduces an additional «syllabification rule, which moves a pre-sonorant obstruent in words such as Vennemann's variable items (see (15)) into the onset of the syllable beginning with the sonorant. This rule is shown in (37). (37)

Resyllabification C O ΐ - Ί χ χ I I [- son] [+ son] Condition: The two segments in (37) are tautomorphemic

Having established what the basic strategies adopted by Hall and Wiese are, we are now ready to examine some derivations, to see how well they perform. Let me begin with some reasonably straightforward items, the adjectives kindlich ([lcintliç], 'child-like') and kindisch (fkindij], 'childish'). As far as I can ascertain, Wiese would derive these as shown in (38). (Both affixes are attached on level 2 in his framework.)

58

(38)

kind-isch

kind-lich

Χ Χ Χ Χ

X X X X I M I k i n d

I I II k i n d

Underlying

Cycle 1 Syllabification

A

A

C V c c

c ν c c I I I I k i n d

k i n d

Cycle 2 Affixation

A

C V c c + x x I I I I II k i n d

A

C V C C+ XX X I I I I I I I k i n d 1 ι ç

ι J σ

ΑΛ σ

σ

C V c c + vc I I I I II k i n d

σ

(Re-)syllabification

A A

C V C C+C V c I N I I I I k i n d 1 ι Ç

ι /

Postlexical σ

σ

A A

Final Devoicing

C V C C+ C V c I N I I I I k ι η t

1 ι ç

[•kindij] [•kintliç]

Output

59

There can be no doubt that Wiese's framework can successfully derive forms such as kindisch and kindlich. Let us compare the derivation in (38) with the equivalent in Hall's approach, which is shown in (39). (39)

kind-isch

kind-lich

XXXX I I I I k i n d

X XXX I I I I k i n d

Underlying

Cycle 1 a

Ο I I X

Ν t . X

Syllabification

σ

• ' .

*

C κ . \ X X

Ο Ν ι I III χ χ

k i n d

C κ ". χ χ

k i n d

Cycle 2 a Λ \ ΟΝC I I κ ΧΧΧΧ

Affixation

σ / Κ Ο Ν C XX

I I κ Χ ΧΧΧ

XXX

k ι η d] ι J

k ι η d] 1 ι ç

σ / Κ ΟΝC

σ : Ν

ΧΧΧΧ

XX

σ / Κ ΟΝ C Μ Ν Χ ΧΧΧ

ι ικ :

k ι η d] i J

σ : Ν : XXX

k ι η d] 1 ι ç

Nucleus and Syllable Placement

60 σ

σ

/ Ν ,Ί Ο Ν C Ο Ν I I Ν Χ Χ Χ Χ

I XX

I M I II k ι η d] ι J

σ / Κ ON C ι ι Ν ΧΧΧΧ

I I II

k i n d ]

σ ,Ί ON : ι XXX

CV Rule

III

1 ι ç Onset

σ

σ

/ Κ χ κ Ο Ν C Ο NC ι ι ι ι ΧΧΧΧ

ι : XX

k ι η d] ι J

a Ο/ ι Χ

K C Ν ι Κ ΧΧΧ

Coda Placement

a l · C. Ο/ Ν ι ι : XXX

k ι η d] 1 ι ç

Postcyclic

[•kindij]

Final Devoicing

σ A \ Ο Ν C II Κ ΧΧΧΧ

σ Λ Ο Ν I I XX

k ι η t]

1 ι ç

['kintliç]

\ C I X

Output

Clearly, these relatively straightforward items can be handled without any difficulty by both approaches. A much tougher test is posed by Vennemann's variable items (given in (15)). In (40) and (42) we can see how the two approaches handle both A-speaker and B-speaker pronunciations of the adjective schmuddlig (['Jmud/tliç], 'grimy'). The first derivation (40) illustrates Wiese's approach, which crucially relies on ambisyllabicity in order to generate the correct A-speaker output (shown on the left). Wiese does not spell out how the syllabification of ambisyllabic segments proceeds nor exactly how these are represented underlyingly, which is why I have omitted the early parts of the derivation. We pick it up at the end of the second cycle.

61

(40)

schmufdjlig

schmu[t]lig

Cycle 2

σ

/h\

/A\

c c ν c c c + vc I I I V I II

J mυ

d

1

ι ç

σ

Syllabification

Ά\ Α\

CCVCC +V c

I I I I I

II

J m υd 1

ι

σ

σ

ç

Postlexical Final Devoicing

A\ Λ\

ccvcc +vc I II I I I I

J mυ t ['/mudi ι ç]

1

['/muti ι ç]

ι

ç

Output

The reason for the failure of rule (29) to apply to the A-speaker form in (40) is that the relevant obstruent is ambisyllabic. This ambisyllabicity means that the SD of (29) is not met, since only part of the geminate structure which represents the ambisyllabic obstruent matches the environment specification. This has sometimes been referred to as the inalterability of geminates (e.g. in Hayes 1986), that is, a special property of geminates which means that any phonological process either applies to both parts of the geminate (if its SD allows it do so) or that it is blocked altogether. Hall's equivalent derivations follow in (42) below. Although Hall does make it clear how an underlying 15ml cluster is to be syllabified (by means of the CV Rule and coronal fricative adjunction), I shall leave out the relevant steps in (42), as they are not immediately relevant to the purpose of showing the derivation, which is to see how it captures the differences in the application of FOD between Α-speakers and B-speakers. What Hall needs to do in order to capture this distinction is to state that the Resyllabification Rule (37) is associated with a different condition for B-speakers than it is for Α-speakers. The former do not apply this rule to Vennemann's variable items because they "have extended the domain of the positive and negative syllable structure conditions [...] to Resyllabification as well' (Hall 1992b: 91). That is, for B-speakers, the

62

condition on the application of Resyllabification must be phrased as in (41), so that the rule is blocked in a derivation such as the one in (42). (41)

Resyllabification C O ΐ - Ί χ χ I [+I son] [- son] Condition: The two segments in (41) are tautomorphemic; positive and negative syllable structure conditions [...] hold.

The derivations resulting from the two different forms of Resyllabification in (37) and (41) are shown below. (42)

schmufdjlig

schmuftjlig

XXXXX I I II I / m u d i

XX XXX I II I I J mυ d 1

a • Ν I

a ! Ν I

x x x x x I I I I I J m u d l

XXXXX I I I I I / m u d i

σ / f \ Ο ΝC .'1 I I

σ /T\ O Ν C . Ί Ι Ι

Underlying

Cycle 1

xxxxx / m u d i

xxxxx j mu d 1

Nucleus and Syllable Placement

Coronal Fricative Adjunction

63

Cycle 2

χ I J

/

σ / Κ Ο ΝC I I I / χ χ χ χ χ χ I II I I I m υ d 1] ι ç σ Α \

σ \

Ο Ν C Ν / ΐ ι ι

χχχχχχχ J

:

m υ d 1] ι ç σ

σ Λ \ . Ί ON C Ο Ν /ι ι ι : ι χ χ χ χ χ χ χ ί m υ d 1] ι ç

σ / Κ ΟΝ C I I I χ χ χ χ χ χ χ I I I I I II J mu d 1]ι ç /κ : Ο ΝC Ν /ili : χ χ χ χ χ χ χ I I I I I II ί m υ d1] ι ç σ σ Α \ Α Ο Ν CΟ Ν /ί ι ι : ι χ χ χ χ χ χ χ I I I I II I J m υ d1] ι ç

Affixation

Nucleus and Syllable Placement

CV Rule

Onset σ

a

/ Κ / Κ Ο Ν CΟ Ν C /ί ι ι ι ι : χ χ χ χ χ χ χ

A \ / T v Ο Ν C Ο Ν C /ι ι ι ι ι : χ χ χ χ χ χ χ

m u d 1] ι ç

J m υ d1] ιç

S

Coda Placement

64 Postcyclic a

Resyllabification

a

A\ A\

ΟΝC OΝC

/I I f Ί I I I II I I Μ J mυ d 1] ι ç χ χ χ χ χ χ χ

a

a

/Κ /N

Final Devoicing

Ο ΝC ΟΝC /I I I I I I χ χ χ χ χ χ χ — ['/mudliç]

IIIIII I

J mut

['Jmutliç]

1]ι ç

Output

In (42),finaldevoicing can apply only to the B-speaker form because the A-speaker form has been resyllabified in such a way that the SD of rule (29) is no longer met. The coronal plosive is now part of a branching onset [dl], Again, the correct output has been generated, although some of the assumptions which have had to be made in order to achieve this are peihaps controversial. I shall return to them in the next section (2.4.3.3). In the meantime, it may be worth considering what happens when there is not a single obstruent which needs to be devoiced, but an obstruent cluster. It is logically impossible for more than one obstruent (or any other segment, for that matter) to be adjacent to a syllable boundary at any one time. So, the proponents of a rule which states that obstruents devoice next to a syllable boundary need some kind of device to enable them to deal with entire clusters. Wiese and Hall make very different proposals for this purpose. For Wiese (p. 82), this problem can be solved with the help of the notion of the appendix. He assumes that an appendix is not genuinely part of a syllable, so that its presence does not affect the application of the Devoicing Rule (29) in any way. So, in a word such as Herbst ([heçpst], 'autumn') or Passivs (['pasiifs], 'passive', gen. sg.), the bilabial plosive and the labio-dental fricative are still syllable-final for the purposes of FOD, because die presence of the appendix /st/ and /s/ respectively makes no difference to the syllable structure. Hall (p. 125), by contrast, rejects this approach (for reasons to be discussed in 2.4.3.3) and proposes the Regressive Voicing Assimilation Rule in (43) to handle cases of Ais kind.

65

(43)

Regressive Voicing Assimilation (RVA) [ - son]

ROOT

I •

·

LARYNGEAL

[- voice] The RVA rule (43) spreads voicelessness from right to left, so that, in a word such as Jagd fljaikt], "hunt*), the final segment (/d/) would be devoiced by the application of the syllable-final devoicing rule (29). It would then spread its [ - voice] specification to the left, which would automatically trigger the delinking of the feature [+ voice] from the underlyingly voiced velar plosive in Jagd.

2.4.3.3 Wiese 1988 and Hall 1992b: some potential problems In this section I would like to consider some potential problems with Hall's and Wiese's accounts. I am not the first to notice difficulties, and a number of perceptive comments on both Wiese's and Hall's work in this context have been made by Yu (1992b: 169ff. 195fF.) and on Wiese's by Hall (1992b: 125). There is no point in rehashing here everything these two authors have said. Instead, I shall summarise one or two of their points which appear particularly relevant to the present discussion and add observations of my own which concern issues not yet raised, at least not in published work. I shall begin with specifics and broaden the scope of my comments later. First of all, I still owe the reader the explanation for Hall's decision not to adopt Wiese's approach to dealing with the devoicing of obstruent sequences (as opposed to individual obstruents). Hall (1992b: 125) points out that, although Wiese is in a position to handle words such as Herbst and Passivs, these do not actually involve the devoicing of more than one obstruent, as the final /st/ and /s/ are underlyingly voiceless. It is where devoicing applies to a sequence of obstruents that Wiese's approach is unable to generate the correct output. Consider the noun Jagd again. If the /d/ is indeed an appendix and as such is not strictly speaking part of the syllable (Wiese 1988: 82), then it cannot be subject to rule (29) either. Wiese seems to have two options, neither of which leads to the desired result. Either the appendix is a proper part of the syllable, which would, incorrectly, prevent /g/ from being devoiced in Jagd, or it is not, which would mean that the Idi, again incorrectly, fails to undergo devoicing. It is for this reason that Hall resorts to the RVA rule (43). This rule, however, is not without its problems either. Hall (1992b: 125) simply assumes that the spreading of [- voice] automatically triggers the delinking of a previously present [+ voice] specification. This is an approach which has become

66

increasingly controversial over the past years. Rice & Avery (1990), following Mascaró (1987a) and Piggott (1988), argue that 'a feature or node can spread only to an empty position' (p. 429). Harris (1994b) also makes the point that spreading-cum-delinking accounts such as this can only really be justified if it can be shown that both operations are independently motivated. This is not the case in Hall's analysis, as he does not interpret devoicing as the delinking of a [+ voice] feature alone. A further weakness of both Hall's and Wiese's analyses appears to be that both of them have to accept that their negative SSCs are systematically violated by words which are structurally similar to Vennemann's variable items (see (15)). Both have SSCs which explicitly exclude /ti/ and /dl/ from the inventory of well-formed syllable onsets and yet the derivations in (40) and (42) crucially depend on their existence. Wiese (1988: 93f.) may be at pains to point out the minor importance of SSCs, but these facts nevertheless raise the question of whether SSCs are worth having and, if not, why both Wiese and Hall find it necessary to employ them after all. This seems to be a somewhat contradictory stance. In any case, the status of SSCs is perhaps somewhat doubtful, but it seems reasonably uncontroversial that the insight captured in Vennemann's (1972, 1988) Law of Initials (22) is a useful one. However, it is precisely this insight which is ignored by any analysis (including Vennemann's own, e.g. Vennemann 1968, 1972, 1978) which crucially depends on syllabifications which assign the status of syllable onset to sequences such as [bn dn di ti zl]. Another possible problem which I would like to draw attention to is that Wiese (1988) needs to invoke the concept of ambisyllabicity. As Harris (1994a: 200) points out, this device is controversial, to say the least. One reason for this is that it requires improper bracketing. Other arguments against it can be found in Selkirk 1982, for example. The general trend certainly seems to be away from ambisyllabicity, as many recent analyses eschew it. Hall (1992b: 49ff.), following a number of other researchers, including Selkirk (1982) and Giegerich (1985), claims that 'ambisyllabicity is simply not relevant for the phonology' (p. 49). A potential weakness which, unlike the issue of ambisyllabicity, is shared by both Wiese's and Hall's accounts is that they can only accommodate the full range of German syllables by allowing for stray segment adjunction (Hall) or by recognising word-initial and word-final appendices (Wiese). Clearly, these devices capture some of the properties of coronal edge segments, but, as we have seen, they can create additional problems (such as how to deal with Jagd for Wiese), quite apart from the fact that they stand completely outside any of the generalisations about (German) syllable structure which are captured in the sonority hierarchies and syllable structure assignment principles/algorithms. It is clear that any account which would capture the exceptional behaviour of coronal edge segments and provide a way of integrating them into the general syllable structure is preferable to the ones proposed by Wiese and Hall. As we shall see in the following chapters, Government Phonology makes this kind of approach possible.

67

Finally, we need to consider the issue of resyllabification. Resyllabification is an extremely powerful device which, in the derivations discussed in 2.4.3.2, takes segment out of apparently well-formed syllables in order to form others, which contain clusters of doubtfiil status. Or, to put it in Hall's (1992b: 89) words, 'resyllabification is capable of generating virtually any word-medial syllable-initial cluster composed of an obstruent and a sonorant'. This seems less than desirable, as it opens the floodgates to systematic overgeneration. In any case, the null hypothesis for any sound theory must be that structures (e.g. syllable-structures), once established, cannot be undone by later operations. I shall have a little more to say on this point in 3.2.1.

2.4.4 FOD applies to segments in a coda 2.4.4.1 Introduction In this section I shall consider a closely related approach to FOD, which, however, has to be distinguished from the one described in the preceding section (2.4.3). This approach makes the claim that FOD applies to obstruents which are syllabified into a syllable coda. It has been quite widely adopted in recent work, including Vennemann 1978, 1982a, Mascaré 1987a, Hall 1989a, Goldsmith 1989, 1990, Rubach 1990, Yu 1992a, b, Walther & Wiese 1993 and Walther 1993. To see what distinguishes the syllable-final and the coda analyses from each other, let us compare the two rules, shown in (44a) and (44b). (The rule in (44b), incidentally, represents an amalgamation of Rubach's (1990) and Yu's (1992b: 167) rules.) (44)

σ

I coda

I a.

[- son]

[- voice] /

]0

b.

[- son]

[- voiced] /

It is clear from (44a) that the SD of the rule is only met by an obstruent which is immediately adjacent to the syllable boundary. In principle, this could cause problems with forms where the syllable ends in more than one obstruent, if not only the right-most obstruent needs to undergo FOD. Take the by now familiar example of Jagd again. If one decided, as Wiese (1988) does, to apply the Final Devoicing ride postlexically and, unlike Wiese, did not allow for word-final appendices, the rule would, counterfactually, generate *[ja:gt] instead of [ja:kt]. Authors who employ this particular rule are well aware of this potential problem and adopt strategies such as the use of appendices (Wiese 1988) or regressive voicing assimilation rules (Hall 1992b). Under the coda analysis, such devices would become redundant, since, if both obstruents are dominated by the coda node it should be possible to apply the relevant rule to both of them.

68

However, this may not be entirely successful either, since, as Hall (1992b: 126) observes, the application of the rule in (44b) to more than one segment dominated by a coda node would be in conflict with Hayes's (1986) Linking Constraint, which states that 'association lines in structural descriptions are interpreted as exhaustive' (p. 331). By this is meant that a rule applies to an autosegment etc. only if it has the same number of linkages as specified in the rule. Hall argues that since the rule in (44b) has only one association line, any input which has more association lines fails to meet the SD of the rule. In other words, instead of [ja:kt], the rule in (44b) would generate *[ja:gd] due to its application being blocked altogether. Not all researchers would agree with Hall's view, though, that the Linking Constraint invariably rules out the coda approach to FOD. Walther (1993), for example points to critical comments on the Linking Constraint made in Scobbie 1991 and puts forward a coda-based alternative which does not employ this constraint. In what follows I shall consider just one representative account of the view that it is the coda which must be referred to in order to identify the environment in which FOD applies. This account is Rubach's (1990), which is so detailed and carefully argued that it is probably the representative of this particular approach to FOD which is worthiest of critical discussion. These merits also mean that it can only be done justice by a thorough examination, which is what I shall attempt to present in the next section.

2.4.4.2 Rubach 1990: the analysis Rubach's solution differs from any discussed so far in that it succeeds both in rejecting the sort of onsets which are in conflict with Vennemann's Law of Initials and in capturing Hochlautung pronunciations of Vennemann's variable items. He phrases the coda-based Final Devoicing rule as in (45). (45)

N'

I I

X [ - son] -»• [ - voiced] /

This rule makes the right predictions in the vast majority of cases, provided that it applies postcyclically,33 when all word formation has been completed. As we have seen above, syllabification needs to refer to all segments which eventually form part of the complete word, and the application of FOD in turn depends on the final syllabification of the word in question. The plural of Tag ([ta:k], 'day'), for example, can be syllabified as Ta.ge

33

See Booij & Rubach 1987 for a detailed discussion of the differences between postcyclic and postlexical rules in Lexical Phonology.

69 (['taiga]) only once the final -e has become available, and it is this syllabification with the plosive in the onset of the second syllable which correctly blocks FOD. To formalise the fact that syllabification takes affixes into account at all levels of derivation, Rubach introduces the Syllable Structure Algorithm reproduced in (46) below. (46)

Ν I X

X N-Placement:

I I [- cons] -*• [ - cons] N" (= σ) /1 / Ν

Ν I

/

I

CV Rule:

( X ) X ^ (Χ)

X

Onset Rule:

Attach the prenuclear Xs to N".

Coda Rule: Erect N' between Ν and N" to include all the postnuclear Xs. N-placement and the CV rule are universal, while the onset rule and the coda rule are subject to language-specific constraints, e.g. constraints on permissible onsets. The CV rule is able to resyllabify codas (except, of course, on the first application of the SSA, where codas are not yet available to the CV rule), but, like Hall's (1992b) CV Rule, it can only fill a single pre-nuclear slot. The purpose of the CV rule is, then, to ensure that the universally preferred syllable type CV emerges, rather than, (VC)0(V)0, which initially results from vowel suffixation. The language-specific onset rule, on the other hand, is responsible for generating branching onsets such as gl, bl and pr in German. It can only syllabify free segments, that is, resyllabification by the German Onset rule is blocked. The SSA applies and reapplies throughout the derivation (both lexically and postlexically) after the application of a rule, whenever its SD is met. To tackle Vennemann's variable items, Rubach first of all makes the morphological structures involved precise. He does this by contrasting some of the relevant variable words with others where FOD is obligatory both for Α-speakers and B-speakers. The appropriate list is reproduced in (47).

70

(47)

Voiced obstruent

Voiceless obstruent

Handl+ung 'act' (handel+n 'to act') Ordn+ung 'order' (ordn+en 'regulate') ebn+en 'flatten' (eben 'flat') Begegn+ung 'meeting' (begegn+en 'meet') eign+en 'own' (eign+en 'to own') nebl+ig 'foggy' (Nebel 'fog')

hand+lich Tiandy' Bild+nis 'portrait' Ergeb+nis 'result' Wag+nis *boldness' Zeug+nis 'testimony' glaub+lich "believable'

Like all his predecessors Rubach has doubts about clusters such as dl or gn forming permissible onsets. In fact, he decides that they cannot constitute onsets at all. Instead, and this is the ingenious move which sets his solution apart from all others - he argues that the sonorants involved become syllabic by application of the Sonorant Syllabification rule shown in (48), which turns all extrasyllabic sonorants (marked with an asterisk) into syllable heads. Extrasyllabic sonorants are defined as those sonorants which cannot be syllabified because of a clash with the Sonority Sequencing Generalisation (see (33)). The relevant scale employed by Rubach for determining how the SSG is to be fleshed out reads 'nucleus - liquids - nasals - fricatives - stops' (p. 80). (48)

σ(=Ν") *C -f C [+ son] [+ son]

The derivation of a word such as Handlung would then go like this. (49)

handl

Cycle 1 a ha η d

Syllable Structure Algorithm

71

σ

σ

• A s

I

Sonorant Syllabification

h a η d 1 Λ σ\

σ Λ Syllable Structure Algorithm

h a η d 1 Cycle 2 α / Κ

σ Λ

h a n d l + u n g a / Κ

a Α

σ / Κ

h a n d l

+ u n g

Syllable Structure Algorithm

Postcyclic Final Devoicing [handluq] The postcyclic application of FOD is blocked because the SD of the devoicing rule (45) is not met, as there is no suitable obstruent occupying a coda position. In handlich, by contrast, FOD applies to the coronal plosive, as this is syllabified in the coda of the initial syllable hand throughout the derivation, with the / forming the onset of the suffix -lich. Rubach observes that the / in Handlung, on the other hand, is not resyllabified into the onset of the suffix -ung (which would incorrectly predict devoicing) because it does not occupy a coda, but a nucleus instead. He assumes that only codas can be resyllabified by the SSA (more precisely, by the CV rule). Thefinalstep in Rubach's account of FOD has to do with the status of those sonorants which were made syllabic by the Sonorant Syllabification rule (48). Some of them remain syllabic until the final output of the derivation (e.g. in Handel, 'trade'), while others, such as those in Handlung and Ordnung undergo a Sonorant Desyllabification rule (never explicitly formulated in Rubach's paper) which is ordered after FOD. It appears that this rule deletes the σ-node dominating the affected sonorant. This means that it (and all other segments dominated by the same node) are completely unsyllabified. Because of this the SSA is able to syllabify them freely (i.e. even the former syllable nuclei can become onsets, although /^syllabification of a nucleus into an onset is blocked). In the case of Handlung, for example, the d would become part of the coda of the first syllable, while the I would go into the onset of the suffix.

72

Rubach claims that he is able to do justice to the fact that FOD is syllable-conditioned and that syllabification depends on morphological structure. Cyclic assignment of syllable structure makes it possible for him to eliminate the apparent conflict between the syllable-based and morpheme-based analyses of FOD and to generate correct outputs both in terms of FOD (i.e. Han[d]lung) and in terms of intuitional 'surface syllabification'34 (i.e. Handlung) for the hitherto problematic pronunciations of Aspeakers.

2.4.4.3 Some problems with Rubach's analysis35 In his analysis, Rubach proposes to treat two of his rules as optional, the German Onset rule, which is part of the SSA and responsible for generating well-formed branching onsets, and the Sonorant Desyllabification rule, which applies postlexically and removes the syllable node from syllabic sonorants, thus freeing both the sonorant and all other segments originally dominated by that syllable node. I believe that optional rules have certain basic properties which are uncontroversial. One of these properties is that both their application and their failure to apply will result in well-formed outputs. If, for example, one wished to express the loss of schwa in a word such as Frauen ([frauan] or [firaun], 'women') in terms of a rule of schwa-deletion, such a rule could be an optional rule. Both the form with schwa and the form without it are perfectly grammatical, with the former being perceived as slightly more formal than the latter by native speakers of German. Rubach's so-called optional rules (German Onset and Sonorant Desyllabification), however, deviate from this pattern. Treating them as genuinely optional has disastrous consequences in a considerable number of cases. It can result in surface forms with unsyllabified or incorrectly syllabified segments and, for B-speakers, it makes the wrong predictions with regard to FOD for certain sets of data. To see that his 'optional' rules actually have to apply in some cases, while they need to be blocked in others to generate the correct output, consider the derivation of the verb handeln ([handln],36 'to act').

34

35

36

See Rubach 1990 (footnote 2; p. 92). Rubach's claim that surface syllabifications such as Hand lung, which are produced by Α-speakers when asked to divide the word up into syllables, 'are performance problems' is vigorously refuted by Hall ( 1992b: 100), who makes it clear that there are speakers of German who can produce only forms which indicate genuine syllabification into Hand lung (Vennemann's B-speakers). I agree with Hall that Rubach is incorrect in his assessment of the facts here. See e.g. Hall 1992b (pp. 96ff), Féry 1991 (pp. 80f.) and Giegerich 1992 (p. 147) for additional critical remarks on Rubach's analysis. This transcription, together with several others to follow below, reflects the output generated by Rubach's derivation. As already mentioned, native speakers perceive these pronunciations without schwa as typical of relatively informal speech, whereas those with schwa, e.g. ['handeln], represent careful or formal speech.

73

(50)

h an dl

Cycle 1 a h a η d *1

Syllable Structure Algorithm

σ

σ I h a η d 1

Sonorant Syllabification

σ σ / Ν Λ h a η d 1

Syllable Structure Algorithm

Cycle 2 a

a

A \

A

hand

1+n

σ σ / Κ A \ h a n d 1+n

Syllable Structure Algorithm

Postcyclic Final Devoicing σ A \ h a η d 1+η

Sonorant Desyllabification

a h a η d 1+n

Syllable Structure Algorithm

•[hand] Rubach makes no provision for the Sonorant Syllabification rule (48) to apply /wsfcyclically as well as cyclically. Consequently, it is impossible to syllabify the / and the «in any way, as the SSA, in the absence of a non-consonantal syllable nucleus, is not able to incorporate them into a syllable either. This would mean that they are unlicensed

74

(see e.g. Goldsmith 1990: 123f.) and would be deleted under Stray Erasure (e.g. Itô 1986: chapter 3). The predicted output is then *[hand], an unattested form.37 This undesirable result could have been avoided if the application of the Sonorant Desyllabification rule in (50) had been blocked. A readjustment rule similar to Kloeke's (see 2.4.2.2.2) might have been able to handle this. However, it would have to permit the application of Sonorant Desyllabification for Handlung (see (49) and discussion above) but block it for handeln and Handel, for example, which would make it necessary for it to refer to specific suffixes. Such a readjustment rule may well turn out to be quite difficult to motivate independently. Another effect of generating unsyllabified segments is, of course, that, once one moves beyond a single word, these segments can then be combined with adjacent segments from other words, giving rise to incorrect surface syllabifications. Now, it may be that cross-word resyllabification is blocked, as no explicit statement to the effect that this is allowed is made in Rubach 1990. On the other hand, it is made clear that the SSA applies 'throughout the lexical and the postlexical components whenever its environment is met' (p. 81). So, it may be reasonable to assume that cross-word resyllabification is permitted. To the extent that this is the case, Rubach's analysis makes incorrect predictions for surface syllabifications. Consider the noun phrase der Wandel in seiner Haltung ([de:g vandl in zainn haltuq], 'the change in his attitude'). The first noun and the following preposition (Wandel in) are of particular interest here, so the simplified derivation in (51) will take only these forms into account. Syntactic bracketing is only partial. (51)

vandl

Cycle 1 σ

v a η d *1 N] σ

Syllable Structure Algorithm

σ

/ f K I v a η d 1N] σ Λ \ Λ ν a n d 1Ν]

Sonorant Syllabification

σ

37

Syllable Structure Algorithm

It might be possible to salvage derivations such as (50) by allowing Sonorant Syllabification to apply postcyclically as well cyclically (after Sonorant Desyllabification). However, for the SD of Sonorant Syllabification to be met, the / would have to be marked as extrasyllabic. It is not entirely clear from Rubach's paper exactly how extrasyllabicity is assigned, but it is conceivable that this could be done by the SSA on its application after Sonorant Desyllabification.

75

Postcyclic Final Devoicing

v'adii d 1 N]

Sonorant Desyllabification

σ χΛΝ. ν a η d 1 N]

Syllable Structure Algorithm

Postlexical a ν a η d

1

σ Λ ν] [pp [ρ 1 η

σ van

d 1 Ν] [ρρ [ρ ι η

Syllable Structure Algorithm

*f['vandlin] The final output of the derivation implies that Wandel in ([vandl in]) is syllabified as [vandlin], with die lateral occupying the onset of in?% German speakers intuitively reject this syllabification.39 As mentioned in 2.4.4.2, Rubach considers native speaker intuitions about syllable divisions important enough to feel the need for phonological representations to reflect them accurately (cf. Handlung, for example). Under this assumption, the output of derivation (51) is simply wrong. However, it is not entirely clear whether native speaker intuitions about the precise location of syllable boundaries are necessarily as reliable40 as Rubach believes them to be. Take, for example, word-internal si-sequences in German, e.g. in fester ([feste], "harder"). Speakers are notoriously unsure about whether to treat this as fes. ter or fe. s ter.

38

39

40

See also Giegerich (1987), who makes the point that, as a matter of principle, no syllabification should be allowed to take place postlexically, as incorrect syllabifications across word boundaries (such as the one in my example) would necessarily arise. In fact, I would argue that thefirstonset in in isfilledby a glottal stop, which of course, would make it impossible for any other segment to occupy that position. Rubach, however, does not consider glottal stop as phonologically relevant and, hence, omits it from his derivations. It is for this reason that I have not included it in my transcription either, although I believe that it does have a place there. I will have more to say about glottal stop in 4.3. I am grateful to John Harris for challenging my own assumptions in this respect.

76

Native speakers also tend to have difficulties where a segment sequence which is perceived as a potential branching onset is involved, especially if there are no unambiguous clues as to the morphological structure of the word. Hutrand (['hu:trant], •brim (of a hat)'), for example, is no problem, as both Hut ([hu:t], "hat') and Rand ([rant], 'edge') are familiar words, and it is clear to any native speaker what the meaning of the compound must be. The fr-cluster is therefore not treated as a branching onset. Stegreif (['Jte:kraif|, as in aus dem Stegreif, 'off the cuff), on the other hand, is interpreted both as Ste. greif and Steg, reif as speakers usually do not know what its morphological structure actually is. Those who opt for the former syllabification frequently misspell it Stehgreif(['/te:gra if]) in the mistaken belief that it is derived from stehen (['Jteisn], 'to stand') and some form of greifen (['graifsn], 'to grasp'). The pronunciation ['Jte:kraif] is the one found in the authoritative DUDEN pronouncing dictionary (Mangold et al. 1990) and corresponds to the syllabification Stegreif with a devoiced velar plosive. This reflects the fact that the word is actually a compound of Steg ([Jte:k], "bridge') and Reif ([ra if], 'ring'), which is derived from MHG stegereif (stirrup'). Similar problems with syllable divisions arise when it comes to some of Vennemann's variable items, e.g. eignen, which is variously interpreted as eignen or eignen. Rubach himself cites further examples of speakers' judgments being contradictory. One, Handlung, has already been discussed. The adjective neblig (['ne:bliç], 'foggy') is a similar case. It retains a voiced [b] in the Hochlautung (suggesting that the orthographic b is part of a branching onset, as a syllable-final b would have undergone FOD), but is nevertheless syllabified as neb.lig (or sometimes ne.blig) by the same speakers. In my experience, for these sort of words there is not just variation between speakers, but the same speaker will come up with different judgments on different occasions. In any case, it is not at all clear whether the sort of competence which is responsible for unmonitored phonological behaviour is necessarily the same as that employed when metalinguistic tasks (such as syllabification and other labelling tasks) are being performed. As Schnitzer (1972: 92) observes, there may be 'a fundamental neuropsychological division between the ability to perform ordinary language tasks (production and perception) and the ability to perform all the other kinds of linguistic tasks'. It seems reasonable to conclude from this that one cannot necessarily expect the same results from both. In short, it appears that there are good arguments against considering native speaker judgments about syllable divisions as an absolute yardstick for phonological representations. In the light of the facts just presented it seems surprising that Rubach finds it necessary to construct his derivations in such a way that they both generate the correct output for unmonitored pronunciations and mirror introspective judgments about syllable boundaries, although the latter are sometimes contradictory. The most obvious way to resolve apparent contradictions between the two (as in the word Handlungen, for example) would have been to model one's representations and derivations on 'normal' (and consistent) speech and to take no notice of intuitions elicited by explicit questioning. Rubach's determination to combine the two, however, occasionally forces him to syllabify the same segment first in one way, then resyllabify it in another and, finally,

77

resyllabify it yet again, in exactly the same way as on the very first pass. This is illustrated for the [d] in Handlung in (52). (52)

handl

Cycle 1 a ha

nd»l σ

Syllable Structure Algorithm

a I

h a n d l

Sonorant Syllabification

σ σ / Κ Λ h a n d l

Syllable Structure Algorithm

Cycle 2 a A \ ha

a A nd

1 + ung

σ σ / Κ Λ ha nd 1

σ Α \ + u n g

Syllable Structure Algorithm

Postcyclic Final Devoicing σ A \ ha

a Λ \ η d 1+

σ ha

ung

Sonorant Desyllabification

σ η d1 +

['handluq]

ung

Syllable Structure Algorithm

78

The d isfirstsyllabified into the rhyme of the first syllable, then it becomes the onset of the second syllable and, finally, it returns to its original position in the coda of the first syllable. In other words, the relevant obstruent is side-lined during the crucial part of the derivation, where FOD applies, and after the application of FOD is put back on track. The two successive resyllabifications of the d which can be reduced to the basic formula A -*· Β A make (52) a Duke of York derivation by Pullum's (1976: 83) definition. Although Pullum himself refrainsfromrejecting the Duke of York gambit as a matter of principle (p. 100), the fact remains that this strategy is viewed with considerable suspicion by a large number of linguists (several of whom are quoted in Pullum 1976), including myself. By arguing that the phonology ought to concentrate on what speakers say without being voy much aware of it rather than on what they think they say (or, more precisely, how what they think they say can be divided up)411 am actually weakening that part of my own argument against Rubach's optional rules which makes reference to incorrect surface syllabifications (as put forward on the basis of derivation (51)). This does not, however, detract in any way from the rather more important point I made about some outputs containing unsyllabified segments (see derivation (50)). Moreover, there is a further argument against Rubach's 'optional' rules which is not connected with intuitional surface syllabifications. The point here is that treating these rules as genuinely optional can lead to incorrect outputs not just for syllabification, but for FOD itself. Before I can show how this comes about, however, I first need to extend Rubach's analysis to Northern Standard German, that is, the dialect spoken by Vennemann's B-speakers. Recall that B-speakers, unlike Α-speakers (whose speech Rubach's analysis is intended to account for), systematically apply FOD in words such as eignen, regnet, ebnen, Ordnung, Adler, edler, Handlung and so on. Now, one could argue that extending Rubach's account to apply it to a dialect other than that for which it was designed is not a legitimate move. On the other hand, it has been demonstrated by Vennemann (1968,1972) that comparing the two dialects (Hochlautung and NSG) is particularly useful for a study of FOD.42 Besides, data from both dialects were available to Rubach when developing this analysis (as witnessed by

41

42

I believe lhat it is perfectly legitimate for a phonological analysis to conflict with what Rubach calls intuitional surface syllabifications, provided, of course, that it captures unmonitored speech. A good theory which can help us gain really useful insights, however, ought to be able to say why such a conflict arises in some cases, but not in others. The Government Phonology analysis, as I will show in chapter 4, is in fact in a position to do that. Some authors (e.g. Kloeke (1982a: 133), Giegerich (1992), Yu (1992b: 177) and Richard Wiese (p.c.)) assume that what I have been referring to as B-speaker forms are, for many speakers, colloquial pronunciations of words which would otherwise be realised as A-speaker forms. I accept this view, but I would still claim that there are speakers (especially in southern Germany) who can produce only A-speaker forms, as well as others (in northern Germany) for whom the production of A-speaker forms requires a conscious effort to conform to Hochlautung. Given this overlap between B-speakers and Α-speakers, it seems all the more important for a successful account of FOD to accommodate both.

79

Rubach's reference to Vennemann 1972, which contains data from NSG as well as Hochlautung). If it turns out that NSG is not at all amenable to Rubach's analysis, then that would be a clear drawback, especially when one considers that speakers from both speech communities exhibit identical usage of FOD, except for the set of words to which Vennemann (1968) refers as 'controversial items'. The Government Phonology analysis to be presented in chapters 3 and 4 accounts for most forms in both dialects and, if it turns out that Rubach's analysis can only handle Hochlautung, the Government Phonology approach would be at a distinct advantage due to its greater generality. I believe, therefore, that an investigation of how Rubach's analysis stands up to NSG data is justified. In order to account for the difference between Hochlautung and NSG in the application of FOD to the variable items, one could either argue that this is due to some difference in morphological structure (e.g. Α-speakers treat the relevant words as morphologically simple, while B-speakers consider them as complex) or that there is a difference in the order in which Α-speakers and B-speakers apply the relevant rules. As far as the items listed in (15) are concerned, there is no evidence to support the claim that the two groups make non-identical judgments about the morphological structure of these words (but see 4.4.2.3 for a more detailed discussion of these, as well as others, where there may be such differences). As argued in 2.4.2.1.2.1, the sort of morphemes (specifically roots) which would be needed to account for differences between A- and B-speakers' pronunciations in terms of morphological structure cannot be motivated. It seems, then, that the difference in FOD behaviour will have to be attributed to a difference in rule order, that is, the rule of Final Devoicing (45) has to be moved from its position in the derivation. On the face of it, it looks as if this can be done very easily. All that needs to be said is that Final Devoicing precedes Sonorant Desyllabification for Α-speakers, while the two rules apply in the reverse order for B-speakers. To see that this simple reordering operation can make the correct predictions, consider the derivation of Handlung (for a B-speaker) in (53). (53)

handl

Cycle 1 a h a η d *1 σ h a n d l

Syllable Structure Algorithm

σ

I

Sonorant Syllabification

80

/ Κ Λ h a η d 1

Syllable Structure Algorithm

Cycle 2 A \ ha

Λ nd

1+ u

ng

σ σ σ / Κ Λ / Κ ha nd 1+ u ng

Syllable Structure Algorithm

Postcyclic a

a

A\

A\

h a n d l σ / I V *

+

u n g

Sonorant Desyllabification

σ ^

/

N Syllable Structure Algorithm

ha

η dl + un

g Final Devoicing

[t] ['hantluq] The reversal of Final Devoicing and Sonorant Desyllabification works well in (53), but it is not watertight. The optional status of Sonorant Desyllabification is still a problem. If I had chosen not to apply this rule in (53), the derivation from the second cycle onwards would have looked like this: (54)

Cycle 2 α Λ \ ha

σ Λ η d 1+ ung

a σ a / Κ Λ Λ \ h a n d l + u n g

Syllable Structure Algorithm

81

Postcyclic Sonorant Desyllabification Final Devoicing •['handluq] Final Devoicing would, counterfactually for a B-speaker, have been blocked. This is another piece of evidence against treating Sonorant Desyllabification as optional. Evidence against the optional status of the German Onset rule (part of the SSA (46)) follows anon. Consider the derivation of Rudrer ([tuidre], 'rower'), a form of the agent noun related to the verb rudern ([ ru:dßn], 'to row'). The derivation for a B-speaker is shown in (55). (55)

ru:dr

Cycle 1

r u: d *r a / K r u: d σ

Syllable Structure Algorithm a I r

Sonorant Syllabification

σ

Λ Λ r u: d r

Syllable Structure Algorithm

Cycle 2 o

a

a

A A

I r+r

r u: d

(according to Rubach (1990:90), by a rule other than Sonorant Syllabification) Syllable Structure Algorithm

82

Postcyclic σ

o

Λ

r u: d r + r σ

Sonorant Desyllabification

σ

/K

rr Hu:. Ad

A

4- r r43 + r

[t]

Syllable Structure Algorithm Final Devoicing

•['ruitrar]44 In this derivation, I have chosen to apply the Sonorant Desyllabification rule, but not the German Onset rule. The output, * [nutre], is incorrect, however. The word is pronounced ['ru:dre] by B-speakers and Α-speakers alike.45 The incorrect output in (55) raises the question of whether we are simply dealing with a lexical exception or whether Rubach's account misses a phonologically significant generalisation and, as a result of this, systematically overgenerates. Let me consider the former possibility first. If what we are faced with is nothing more than a lexical exception, then we would expect to find phonologically similar words which do undergo the process in question, specifically, which are derived correctly, even if the German Onset rule fails to apply. Some words which are similar to Rudrer in the sense required here (i.e. an FOD obstruent is followed by r in a potential onset position) are listed in (5Ó).46

43

44

45

46

Incidentally, there is some possibility that this r cannot be syllabified into the onset of the final syllable, as this would give rise to a syllable with an onset which is as sonorous as the nucleus. This could well be interpreted as a violation of the Sonority Sequencing Generalisation. If this turned out to be the case, thai the r would remain unsyllabified in the final output of the derivation, a definite problem for Rubach, but not immediately relevant to the particular point I am making here. The output of the complete derivation would, of course, have been *['ru:tre]. I have omitted the steps which convert the /ar/ sequence into [B], for reasons of exposition. Hall (1992b: 94) provides additional data supporting this claim, which is based on my own informant work with B-speakers. Readers who are familiar with German may find the spellings used here somewhat surprising. I can assure these readers that, according to the DUDEN Großes Wörterbuch der deutschen Sprache (Drosdowski et al. 1976-1981), they - like Rudrer - are legitimate alternatives to the more common spellings with an e between the obstruent and the r. I have chosen the e-less forms to provide an environment where at least the potential for the application of FOD exists. Obviously, where the e is pronounced (as schwa), FOD cannot apply. Words such as faserig have no e-less alternative spellings, whereas unsrige does. Quite often, though, this orthographic e is not realised.

83

Zaubrer Plaudrer Wandrer Zögrer

['tsaubre] fplaudre] [Vandre] ['ts0:gre]

'wizard' 'so. who chats' 'rambler' 'so. who hesitates'

unsrige fas(e)rig

[Tunzriga] ['fa:zriç]

'ours' 'stringy'

The (comparatively small number of) words which are similar to Rudrer all behave identically to Rudrer, that is, even B-speakers cannot devoice here. Clearly, then, a phonological generalisation is being missed. Looking at (56a) only, one may get the impression that die problem can be solved by making the German Onset rule obligatory. After all, the clusters involved in (56a) are all well-formed onsets in German. This solution, however, is disqualified as soon as (56b) is taken into account as well. From Rubach's discussion of permissible and impermissible onsets in German it can be safely deduced that he would consider the cluster [zr] as belonging to the latter category, so that the German Onset rule would necessarily be blocked in the derivation of unsrige, for example. What dus suggests is that the special properties of words such as those in (56) (i.e. the absence of FOD) have nothing to do with the opposition of well-formed vs. illformed German onsets, but with the special characteristics of r, as opposed to / and η (or m). I will return to this point in 4.5.3. What matters at this stage is the insight that a generalisation concerning r is being missed. Incidentally, a closer look at the derivation of unsrige (for a B-speaker) sharpens a point I made earlier about the occasional incompatibility of unmonitored phonological behaviour and native speaker judgments about syllabification. Consider the derivation in (57). (57)

u n z *r

Cycle 1 σ / Κ u η ζ

*r

σ

σ

Syllable Structure Algorithm

/ Κ Ir u n z

Sonorant Syllabification

σ σ Λ Λ u η ζ r

Syllable Structure Algorithm

84

Cycle 2 a

a

Α

A

u η ζ r + i g σ Α Λσ Λσ

u η ζ r +i g

Syllable Structure Algorithm

Cycle 3 σ

o

a

A A A unz a

r+i a

g+a

a

a

A A IA

unz

r+ig +3

Syllable Structure Algorithm

Postcyclic σ

σ α

A I A u η ζ r + i g+a a / N

a A

unzr+i [s] *["unsrig3]

Sonorant Desyllabification

a A g+3

Syllable Structure Algorithm Final Devoicing

The application of the optional rule of Sonorant Desyllabification in this derivation proves that even the Duke of York gambit identified earlier is not always able to allow Rubach to capture both the unmonitored phonological behaviour of speakers and their judgments on surface syllabifications. After all, it is the application of this rule which allows ItJ to become part of the coda of the, first syllable, thus triggering FOD and generating an incorrect output. At the same time, the application of the rule is vital for producing the correct surface syllabification. In other words, the application of the rule both makes and breaks the derivation. The point is that, without Sonorant Desyllabification, the output would have been a word with four syllables. Native speakers of German, however, with a high degree of consistency, feel that unsrige consists of three syllables (unlike its variant form unserige,

85

which is considered to have four syllables). So, for Rubach's requirement that the output of a phonological derivation has to match intuitional surface syllabifications to be met, Sonorant Desyllabification has to apply. On p. 74 I said that unwanted applications of Sonorant Desyllabification could perhaps be blocked by a readjustment rule. Now it is clear that constraining the application of Sonorant Desyllabification in order to ensure the correct output for FOD would not only require a rule which might be hard to motivate but the jettisoning of one of Rubach's avowed aims, to capture surface syllabifications. Blocking Sonorant Desyllabification here would mean generating the wrong number of syllables (four instead of three). Clearly, then, it is impossible to do both the things Rubach sets out to achieve, at least as far as speakers of NSG are concerned. Let me return to my earlier claim that cases such as Rudrer illustrate the point that Rubach's analysis fails to capture certain special characteristics of r 4 7 This problem arises for B-speakers only. Considering that Rubach's analysis was developed for A-speakers rather than B-speakers, it may, then, not be all that serious. What still remains and is probably more serious, however, is the fact that Rubach's solution lacks generality. It is not possible to extend it to a dialect which is very similar to the one for which it was developed, at least not without creating further problems. Let me summarise, then, what has been said in this section. I have shown that Rubach's analysis suffers from several more or less serious shortcomings. The optional rules of German Onset and Sonorant Desyllabification have turned out to be anything but optional. For Α-speakers, they can generate unsyllabified (and hence prosodically unlicensed) segments in the final output of a derivation, and, if Rubach's approach is extended to B-speakers, even incorrect predictions for FOD itself can arise. It has become apparent that part of the problem lies in Rubach's attempt to combine a phonologically correct output with one that matches native speaker judgments about surface syllabification, despite the fact that the latter are frequently inconsistent. If he had allowed his derivations to reflect only the former, then at least some measures could have been taken to restrict the application of his 'optional' rules.

2.4.5 The feature [voice] is licensed only in a syllable onset: Lombardi's (1991) analysis and its problems To conclude this chapter, I shall very briefly discuss an insightful analysis which turns the accounts dealt with in 2.4.3. and 2.4.4 on their heads, as it were, by referring to the complementary environment. Instead of singling out syllable boundaries or codas and

47

As pointed out to me by Richard Wiese, these special properties extend well beyond the application of FOD for B-speakers. The fact that the schwa-ful form Ruderer exists alongside the schwa-less form Rudrer, whereas schwa-ful forms such as *Segeler ([zeigate], 'sailor') and *Ordener ([ PoçdanB], 'usher") are ungrammatical, provides further evidence in favour of assuming that r is no ordinary sonorant. I shall have more to say on these matters in 4.4.2.6 and 4.5.3.

86

invoking regressive voicing assimilation and appendices or dodging the effects of the Linking Constraint, Lombardi's (1991) approach is based on a single positive wellformedness constraint, the Voice Constraint (58), which stipulates that '[voice] is only licensed in an obstruent if it stands before a [+ son] segment in the same syllable' (p. 39). In other words, this privative feature is licensed only in what is probably best understood as an onset. The Voice Constraint is formalised as in (58), where [± son] has its usual interpretation. LAR denotes the LARYNGEAL node, which is licensed under exactly the same circumstances as [voice], so that a voiceless obstruent is understood to have neither. (58)

σ [- son] [+ son] I LAR·

I

[voice] Spelling out (58) in more detail, Lombardi states that voiced obstruents in languages such as German can occur only in one of the two configurations shown in (59). (59)

a.

σ / \ [- son] V

I

[voice]

[- son] [+ son] V

I

[voice]

It is clear from the phrasing of the Voice Constraint (58) that how well this approach performs depends entirely on how appropriate the syllabification mechanisms are which accompany it. Unfortunately, the focus of Lombardi 1991 is such that no explicit statements about syllabification principles or algorithms are made. There would be no point in speculating about how syllable structure would be assigned and what possible problems could arise from the particular mechanism employed (e.g. resyllabification). Instead, I shall briefly examine the actual syllable structures which would need to be arrived at in order for (58) to make the right predictions. In the light of the derivations discussed in 2.4.3.2 and 2.4.4.2, we can probably be quite confident that reasonably straightforward items could be handled without any difficulty. Kindisch and kindlich, for example, would be accounted for by (59a). The real test, as always, comes from Vennemann's variable items. Lombardi herself confronts her proposals with appropriate data For this purpose, she uses Rubach's list in (47), which is repeated here as (60) for convenience. The items in the right-hand column pose no problems at all for her, as the devoiced obstruents occur syllable-finally and would, therefore, be expected to devoice.

87

(60)

Voiced obstruent

Voiceless obstruent

Handl+ung 'act' (handel+n 'to act') Ordn+ung 'order' (ordn+en 'regulate') ebn+en 'flatten' (eben 'flat') Begegn+ung 'meeting' (begegn+en 'meet') eign+en 'own' (eign+en 'to own') nebl+ig 'foggy' (Nebel 'fog')

hand+lich "handy' Bild+nis 'portrait' Ergeb+nis 'result' Wag+nis "boldness' Zeug+nis 'testimony' glaub+lich 'believable'

What she suggests for the forms in the left-hand column is the following. Considering the fact that the roots of the items in this column can also surface with an overt vowel ([9]) between the obstruent which fails to devoice and the following sonorant, it seems reasonable to conclude that an underlying vowel slot is present here. The relevant forms are listed in (61), which is a modified version of (19). (61)

Handel Orden eben gegen eigen Nebel

[handal] ['îoçdan] ['?e:ban] ['gergan] [?aigan] ['neibal]

'trade' '(religious) order' 'level' 'towards; against' 'owned by; peculiar to' 'fog'

In order to derive the forms in the left-hand column of (60), all that has to be said, then, is that the Voice Constraint (58) 'is only active at an early level of syllabification in German' (p. 68), so that, by the time the vowel slot is deleted, this constraint no longer applies. The derivation of eignen based on this assumption would then go as in (62). (62)

eigVn+en — eignen

Underlying Devoicing (blocked, since [g] is syllable-initial) Vowel Deletion

As far as I can see, Lombardi's proposals also work for B-speakers (although she makes no reference to their pronunciations). The correct forms would be generated by simply removing any stipulation to the effect that the Voice Constraint (58) shuts off at some stage in the derivation. On the one hand it seems clear that Lombardi's very simple approach avoids some of the pitfalls associated with the accounts discussed in earlier sections of the present

88

chapter. On the other hand, difficulties may still arise from it. Two potentially troublesome areas spring to mind straight away. Firstly, arguing that an underlying vowel slot is present and that this is deleted is a move which has some disadvantages. For one thing, it requires resyllabification, as the consonant which formed the onset of the syllable containing the underlying vowel slot does not remain an onset. Instead, it is incorporated into the coda of the preceding syllable. The general objections which can be voiced regarding resyllabification have already been aired in 2.4.3.3, so I shall not reiterate them here. There is a more specific point to be made in the context of the underlying vowel slot in any case. The fact that deletion is required is very much out of step with recent accounts of schwa/zero alternations.48 To my knowledge, the most recent extended account of this phonological event in German to employ wholesale deletion rules is Strauss 1982. In more current work (e.g. Giegerich 1987, 1989, Wiese 1988, Hall 1992b), a clear consensus has been reached in favour of dealing with this by means of epenthesis rules. The reader is referred to the works cited for arguments. The second point I want to raise here has to do with the fact that Lombardi can only account for A-speaker pronunciations of words such as those listed in (61) by stipulating that the Voice Constraint (58) is not active after the earliest stages of the derivation. I am perfectly aware that the shutting off of filters and constraints is standard practice, especially in Lexical Phonology, but this, like resyllabification, is certainly not the position of choice. The null hypothesis is that whatever constrains the well-formedness ofrepresentationsis effective at any stage of the derivation. An analysis which can adopt this strong position and still cover the same empirical ground must, therefore, be preferred. I will propose such an analysis in chapters 3 and 4.

2.5 Conclusion In this chapter I have examined a range of earlier treatments of FOD from 1968 to 1993. I have shown that all of them suffer from more or less serious flaws, most of which originate from the theoretical framework the analysis in question is couched in. Most importantly, in my view, especially the earlier accounts discussed are essentially arbitrary in the way they identify what final devoicing consists in, and they are unable to provide unified and well-motivated answers to the question of where it occurs. In the next chapters, I put forward my own analysis of FOD, in the framework of Government Phonology. This approach, I believe, overcomes many of the problems which beset the earlier solutions examined here and, in doing so, makes explicit most of the reasons for the difficulties arising from previous approaches. 48

In 4.4,1 shall put forward an analysis which, like Lombardi's, relies on the presence of an underlying vowel slot or, to put it more precisely, a nuclear position. I will show that such positions can indeed have a role to play, without any need for their deletion arising.

3. What is Final Obstruent Devoicing? A Government Phonology Approach

3.1 Introduction In this and the following chapter I develop an analysis of FOD in the framework of Government Phonology (henceforth GP). The purpose of these chapters is twofold. Firstly, an analysis couched in a relatively highly constrained framework (pace Coleman 1995) can reveal some facts about the phonological event of FOD which appear to have gone unnoticed in other frameworks. By the same token, some problems for earlier analyses can be resolved in the framework to be presented here. Secondly, confronting a complex phenomenon with a comparatively new theory is bound to raise some issues which indicate that areas of the theory need to be further developed, refined or even modified. In other words, although - or rather, because - my present approach to FOD is within the theory of GP, I would like to take this opportunity to both extol the virtues of GP and to provide constructive criticism of some of its shortcomings. Sometimes, especially where these weaknesses in the theory are not directly linked with what I have to say about FOD, I will simply note that there is a problem, while on other occasions I will try to suggest (partial) solutions, in the hope that, if some of my proposals turn out to be inadequate, they will at least provide a useful point of departure for others. The most important statements of GP policy are made in Kaye et al. 1985 and Kaye et al. 1990 (henceforth KLV 1985 and KLV 1990 respectively). Further important work in this framework is contained in, for example, Charette 1988, 1989, 1990, 1991, 1992, Gussmann & Kaye 1993, Harris 1990, 1992, 1994a, Harris & Kaye 1990, Kaye 1987, 1988, 1990a, 1990b, 1992, 1993, Harris & Lindsey 1995 and Lindsey & Harris 1990. What crucially distinguishes Government Phonology from many other phonological theories is the fact that it uses neither rewrite rules of the type A -> Β / C D nor binary features. The absence of this particular kind of rewrite rules is, of course, a property which GP shares with any full-blown autosegmental theory,1 but, much more importantly, GP, unlike most mainstream theories today, does not employ any rules whatsoever. This makes it akin to constraint-based theories such as Optimality Theory (e.g. Prince & Smolensky 1993), although GP, like recent work in syntax, operates with principles and parameters, rather than violable constraints. The atoms that can be manipulated within the theory are unary primes with realisational autonomy (so-called 'elements'; see Harris & Lindsey 1995) rather than binary features. Binary features are also rejected by sister segmental theories such as Dependency Phonology (e.g. Durand

1

See e.g. van der Hulst & Smith 1982a, Goldsmith 1990 or Kenstowicz 1994 for useful overviews of autosegmental theories.

90

1986b, Anderson & Durand 1987, Anderson & Ewen 1987), Particle Phonology (Schane 1984) and Radical CV Phonology (van der Hulst 1992). I will introduce the details of these and several other aspects of GP as and when they become relevant to my discussion of FOD, but before moving on to this, I would like to present some central notions of GP, together with some 'ground rules' as well as some principles of grammar. In order to avoid losing sight of the true concern of this book, an investigation of FOD and related issues in German, I shall keep the discussion in this section reasonably short. Readers who would like to know more about the arguments which lead to the claims summarised here are urged to consult Harris 1994a (on which much of what follows is based), Kaye 1993, Harris & Lindsey 1995 and Brockhaus 1995.

3.2 Phonological licensing, government, 'ground rules' and some principles of grammar 3.2.1 Phonological licensing and government One of the central tenets of Government Phonology is that licensing is the motor which drives phonology. Every skeletal position within a domain, except for the head, has to be licensed,2 as stated in the Licensing Principle (1) (Kaye 1990b: 306). ( 1)

Licensing Principle All phonological positions save one must be licensed within a domain. The unlicensed position is the head of this domain.

Licensing considerations can be taken to be ultimately responsible for the vast majority of phonological events. Licensing relations are asymmetric binary relations entered into by phonological positions. A position which licenses another position is usually referred to as a licenser (or head), and the 'area' over which a licensing relation extends is known as a licensing domain. Licensing relations can be established at different levels of representation, but they are invariably local, that is, positions which enter into a licensing relation must be adjacent at the relevant level of projection. Government, from which the theory derives its name, is a special, particularly restrictive form of licensing. Governing relations, like licensing relations, are asymmetric binary relations which capture universal phonotactic restrictions. To see what role licensing (including governing) relations have to play in phonological representations we need to consider the phonological constituents provided for in GP. There are three constituents, viz. onset (O), nucleus (N) and rhyme (R). All three may branch, so that the maximal expansions are those shown in (2). 2

This idea, of course, is not unique to GP. Similar suggestions are familiar from earlier work, most notably Itô 1986.

91

(2)

b.

a.

c.

O χ

R

Ν χ

X

X

X

(χ)

X

It is immediately obvious from (2)3 that the relationship between the left-most skeletal position and its sister to the (far)rightdiffers between onset and nucleus on the one hand and the rhyme on the other. Constituents are assumed to be universally left-headed in GP (as indicated by the emboldening), so we can make the following observations. In the case of onset and nucleus, the left-most position is the immediate head of the constituent node, while in the case of the rhyme the corresponding position is the ultimate head of the constituent. This means that the right-most position in O and Ν is a complement of the left-most position, while the right-most position of R is an adjunct. One possible prediction made by this is that the phonotactic restrictions applying within an onset and a nucleus should betighterthan those which apply within a rhyme. This is certainly true, since, as we shall see in the remainder of this book, the rhymal adjunct position can tolerate a wide variety of segmental material, while the onset and nuclear complement positions are much more restricted in this respect. Harris (1994a: 168) proposes to treat branching onsets and nuclei as governing domains, that is, as 'areas' over which a governing relation extends. In his view, the constituent head governs its complement in an onset and in a nucleus. This interpretation is veiy much in line with the body of GP literature. Harris deviates from the view adopted by other GP authors, though, in assuming that no governing relation exists between the head of the rhyme and the rhymal adjunct. This can be justified by considering the range of melodic content which can be tolerated in the coda, which is why I adopt his approach in what follows. Governing relations, being a (special) case of licensing relations, can be established only between positions which are adjacent at the relevant level of projection. As we have seen in (2), positions which enter into (intra-)constituent governing relations are adjacent at the skeletal level. However, not only positions dominated by the same constituent node are adjacent at this level. Positions at the edges of contiguous constituents also display string-adjacency. In principle, then, it may be possible for such positions to enter into governing relations as well. There is general agreement among GP researchers that such governingrelationsdo indeed exist and that they are universally right-headed, although it is less obvious exactly where they can be established. The most conservative strategy is to allow for inter-constituent governing relations only where the most clear-cut phonotactic restrictions can be observed. For Harris (1994a: 168) this means that the only inter-constituent governing relation to be countenanced is that between an onset head and a preceding rhymal adjunct (or 'coda', for short). This is illustrated in (3), where the arrow indicates the direction in which government flows.

3

Whether the Ν appears directly above the skeletal position which constitutes the nuclear head or slightly off-set to the right is of no significance.

0 χ

χ

1

χ

In fact, the Coda Licensing Principle (Kaye 1990b: 311, Charette 1991: 120) restricts the occurrence of rhymal adjuncts to those environments in which they are licensed through being governed by a following onset. I will give a more detailed account of this principle in 4.2.1. Allowing for governing relations to be established exclusively in the configurations shown in (2a), (2b) and (3) has the desirable effect of ensuring that only positions which can never assume the role of constituent head are governees or, to put it differently, constituent heads are never governed in governing relations at the skeletal level.4 We have seen that governing relations are universally left-headed when holding within constituents and universally right-headed when holding between positions in separate constituents. Given that governing relations are a restrictive form of licensing relations, we could assume that these universal statements about directionality also apply to licensing relations in general. Licensing relations can also be established between projections of nuclear heads. Their directionality here is not fixed universally but depends on appropriate parameter settings. What these parameter settings are can usually be determined with reference to phenomena such as 'stress, harmony, vowel reduction and syncope' (Harris 1994a: 157). It is worth noting that, although the directionality of licensing relations at the same level of nuclear projection remains constant within a particular language, the direction in which licensing flows may change from one level to the next. For example, one could argue that, in German, licensing relations at the foot level are left-headed, while at the word level they are right-headed.5 A direct result of the fact that licensing (and hence governing relations) are strictly directional and strictly local is that constituents are maximally binary, as it is logically impossible for both strict directionality and strict locality to be respected in a branching constituent which is anything other than binary.6 At first sight, restricting constituents to maximally two positions may appear to be an approach which is faced with numerous counterexamples, e.g. initial s + consonant sequences such as, say, sir in Italian or English. As shown in KLV 1990 and Kaye 1992, however, the members of this sequence

4

5

6

As we shall see in 4.4.2.2., where a governing relation between nuclear heads is introduced, the reference to the skeletal level is crucial here. Governing relations established at a level of nuclear projection may indeed involve constituent heads being governed. This position is adopted by Yu (1992b: 17), for example. See also Ramers (1992) for a brief survey of the literature on stress in German. See e.g. Kaye 1987 (p. 132), Kaye 1990b (pp. 306f.) or Charette 1991 (pp. 16f.) for the proof.

93

are not syllabified into a single constituent, but two adjacent constituents, as illustrated in(4). 7 (4)

R

Χ

Χ

Χ

X

I I I

s

t

r

To the extent that this analysis can be applied to other languages (and there is no particular reason why this should be impossible), such s + consonant sequences do not constitute counterexamples to the GP claim that constituents are maximally binary. What does seem to constitute a genuine problem for the binarity theorem is the superheavy rhyme structure alluded to in (2c) and spelt out in (5) below. (5)

R

Χ,

X2

Xj

It seems impossible to deny that this rhyme is ternary and is, therefore, ruled out by the strict directionality and locality of constituent licensing. Given that x, is the head of the constituent, it is unable to license x3, since the two are not adjacent. It is for this reason that the majority of GP researchers consider the structure in (5) to be ill-formed. At the same time, one can hardly pretend that words which appear to have precisely this structure do not exist, e.g. the English nouns paint, child and beast and the German nouns Freund ([ftoint], 'friend') and Feind ([faint], 'enemy'). Harris (1994a: 163) argues that, although x3, the right-most position in (5), cannot be licensed by the nuclear head due to its non-adjacency, the position can still receive the necessary licensing from the following onset position, which governs it. So, a word such as beast would have the structure in (6), where the coda position is licensed only by the onset occupied by the coronal plosive.

7

Psycholinguistic evidence supporting this view is provided by an experimental study specifically designed to test the claim that s + consonant clusters are heterosyllabic which was carried out by Treiman et al. (1992).

94

(6) Ο

Ν

It is clear, however, that a coda position which is licensed only by die following onset has a different status with regard to licensing than a corresponding position which is part of a rhyme constituent dominating only two skeletal positions (i.e. a single nuclear position followed by a coda). The latter position would receive a double dose of licensing, as it were, onefromthe nuclear head to its left and one from the onset to its right. For reasons to be spelt out in 4.3 (and discussed in detail in Harris 1992,1994a: 154ff), the differing licensing status of the two coda types can be expected to have a bearing on the segmental material which may be associated with these positions. Specifically, the prediction would be that a doubly licensed coda can contain a broader range of material than a singly licensed one. This prediction appears to be correct in the majority of cases, since a doubly licensed coda may determine its own place of articulation (e.g. acht ([axt], 'eight') or Luft ([luft], 'air') in German and whisk or loft in English), while a singly licensed coda usually has to share the (coronal) place of articulation of the governing onset (e.g. paint and wild in English, and Freund and Feind in German).8 Not only the typical binarity of constituents may seem surprising about GP but also the absence of the syllable and the coda from the inventory of constituents.9 Theoryintemally, a syllable constituent would be anomalous in being the only right-headed constituent, given that each nuclear head licenses the immediately preceding onset head. Theory-externally, there is no conclusive evidence for the existence of such a constituent. Arguments for the existence of the syllable (such as those advanced by Selkirk (1978, 1982) and Nespor & Vogel (1986: 72ff.), for example) can be reinterpreted as arguments for the kind of rhyme we find in GP, together with the concept of nuclear projection. A lack of conclusive evidence in favour can be pointed to in the context of the coda qua constituent as well. Incidentally, the view that there is insufficient evidence for establishing a coda constituent is shared not only by GP researchers but also by authors working within other theoreticalframeworks.They include Giegerich (1992: 168), Wiese (1988: e.g. 82) and Hogg & McCully (1987: 45ff.). Having established that GP allows for three constituents and considered their maximal forms as shown in (2), I would like to look at their minimal forms, too. A nuclear constituent head invariably has some kind of licensing work to do, so that it must always

8

9

I am not aware of any exceptions to this in German, at least none which contain clear-cut coda-onset sequences, while English (RP) does appear to contain several, e.g. daft, laughter, example and sample. See KLV 1990 and Brockhaus 1995 for a more detailed discussion.

95 be understood as the projection of a skeletal point. An onset head, by contrast, may well find itself without licensing responsibilities, so that onsets without a skeletal point are assumed to exist. At die same time, there are arguments (summarised in Brockhaus 1995) for distinguishing between onsets with and without a skeletal point. Such arguments comefromphenomena such as apocope in French, where a definite article such as le or la preceding an onset without a skeletal point loses its vowel and emerges as /'. A noun phrase such as la amie ('the femalefriend')would actually be spelt l'amie, which reflects the realisation [lami]. Exactly what the mechanism involved may be will be discussed on p. 102. All we need to say at this point is that this apocopation of the vowel of the definite article is blocked where the following noun begins with h-aspiré (which is completely inaudible). So, a word such as havre ([avr], "haven'), which, from a phonetic point of view, is just as vowel-initial as amie, patterns with consonant-initial words in not triggering apocope, yielding le havre ([lsavr]). It has been argued (e.g. by Charette (1991: 90), following Vergnaud (1982)) that, where apocope takes place, two nuclei are adjacent and one of them is lost, while, where apocope is blocked, e.g. in the case of h-aspiré, the two nuclei are separated by a skeletal position, projected to an onset. The positionless configuration is shown in (7), where (7a) illustrates the lexical representation of l'amie and (7b) the surface form.10 (7)

a.

O,

N,

02

N,

03

N3

b.

O,

I I X X

I I I X X X

I

I I l a

I I I a m i

I

N2

03

N3

I x

I x

I x

a

I m

I i

x l

I

Forms beginning with h-aspiré, by contrast, are characterised by an onset which does have a skeletal position, so that the two nuclei are not adjacent, as shown in (8). 0 1 X I 1 1

Ν 1 X I 1 3

0 I X

Ν 1 X ι 1 α

/ X I1 ν

0 Ν \ I X X 11 r

It is now clear why we need to allow for a minimal onset to be able to take one of the two forms shown in (9a), while the minimal form of a nucleus must contain a skeletal position, as indicated in (9b).

10

As already mentioned, the mechanism involved in deriving (7b) from (7a) will be spelt out on p. 102.

96

(9)

a.

O

O

I

χ

b.

Ν

I

χ

For licensing relations to hold, the segments associated with the skeletal positions which enter into a licensing relation have to fulfil certain requirements. The details of this will be discussed in the section on segmental representations (3.3), so suffice it to say at this stage that charm and complexity11 were taken to be crucial in this context in most of the GP literature until about 1992. Since then, charm has gradually disappeared as a result of the realisation that, although it contributed to some extent to reducing the range and number of segments generated by the theory, it also made some blatantly false predictions. For example, therigorousapplication of charm theory would entail denying that nasalised low vowels exist. This, of course, is untrue, as witnessed by languages such as French. It is for this reason that I shall have nothing to say about charm. We have seen that a position (except for the head of a domain) can be licensed by being incorporated into a licensing hierarchy through the establishment of licensing (including governing) relations. However, this is not the only way in which a licence can be gained. Domain-fmal empty nuclear positions can also be licensed by parameter setting, a possibility which will be discussed in more detail in 4.2.1. For the time being, let me simply note that these positions are licensed to remain inaudible, although they dominate a skeletal position. Kaye (1992) also argues for a third licensing mechanism, so-called 'magic licensing'.12 This is invoked in the context of s + consonant clusters in order to account for the licensing of the empty nuclear position which licenses the coda occupied by [s] (or [j] in German). The configuration illustrated in (10) would be that found at the left edge of German words such as Stein ([/tain], 'stone').

11

12

See especially KLV 1985, 1990 and Kaye 1990a for discussion of charm and Harris 1990 for a detailed account of complexity. The term 'magic licensing' was chosen 'as a constant reminder that it is pure stipulation in need of explanation' (Kaye 1992: 306). Incidentally, 'magic licensing" is probably not restricted to s + consonant clusters. As Pagoni (1993:116-117) argues, the same mechanism may need to be invoked in order to account for the existence of certain nasal + oral stop clusters in Modem Greek.

97

The licensing of this inaudible nucleus cannot be brought about by any other means, as we shall see in chapter 4. 'Magic licensing' is, as yet, only poorly understood. It is not central to my investigation of FOD, so I shall not discuss it any further. To conclude this subsection, I would like to consider at what stage licensing relations are established. According to KLV (1990: 221), governing relations (or, in the terminology used here, licensing relations, including governing relations) are established at the level of lexical representation, where the level of lexical representation is defined as 'the level at which the stem is attached to accompanying affixes, if any' (KLV 1990: endnote 34). It is not immediately clear to me how exactly this is to be interpreted. To facilitate the discussion of this issue, let us assume that the establishment of licensing relations (and hence constituent structure, because the two are inextricably linked, as we have seen) can be equated with the assignment of syllable structure. This is something of an oversimplification, but it should help readers unfamiliar with GP to conceptualise what is involved here. Given the views most widely adopted in the literature, we have essentially two options then. We could either assume that licensing relations are established on the basis of lexical representations which are largely redundancy-free or that lexical representations already contain syllable structure. Let me discuss each option in turn. A possible argument in favour of building licensing relations onto lexical representations which contain only information which is clearly not predictable would be that redundancy is minimised. Ever since the early 1960s, the period when the first statements on the evaluation metric were made (see Hyman 1975: 103ff.), redundancy has been abhorred by phonologists. The development of Underspecification Theory (see e.g. Archangeli 1984, 1988 and Kiparsky 1982b) is in a sense a consequence of treating a simpler grammar (in terms of countable features) as being more highly valued than a relatively more complex one. In other words, from the point of view of the evaluation metric, a lexicon which contains redundancies is not highly valued. Taking the evaluation metric as an absolute yardstick, we would then have to assume lexical representations such as the one shown in (11), where only unpredictable information is present. (11)

Ν

I I I I

X X X Ρ

I

t

Licensing relations would then be built by the application of universal principles and language-specific parameter settings to convert the lexical representation in (11) into the one in (12), the details of which will be discussed in later sections of this book.

98

(12)

Ο

Ν

Ο

l i l i χ χ I I I

χ

Ρ

ι

Ν χ

t

Alternatively, one could claim that licensing relations are present in lexical representations. This would amount to adopting the templatic approach put forward by, for example, Selkirk (1982) and Itô (1986). Selkirk (1982: 356f.) explicitly makes the point that syllable structure is present in underlying representation: We will not assume that the principles of BSC [basic syllable composition/WGB] 'apply', in the sense that they participate in a phonological derivation, converting a phonological representation. Rather, we think of them as well-formedness conditions on underlying phonological representation, which thus is to be thought of as having syllabic structure.

In this statement, Selkirk puts her finger on the crux of the distinction between the two approaches to syllable structure assignment. What matters is not whether syllable structure is always present in lexical representation or whether it is assigned by algorithm to lexical representations. As pointed out by Harris (1994a: 84), these two views are no more than notational variants. What is really important is whether the assignment of syllable structure is affected by the application of phonological processes proper. Rubach (1990), makes a case for assigning syllable structure throughout the derivation - a strategy which is quite widely adopted (see e.g. Wiese 1988: 83ff„ Hall 1992b: 8 Iff., Yu 1992b: 54ff., Giegerich 1992 for German). The application and re-application of a syllable structure algorithm is considered necessary by most of the authors mentioned so far, but not all of them see resyllabification as a valid option. Saying that syllable structure is subject to change in the course of the derivation, as Rubach (1990), Wiese (1988), Hall (1992b) and Yu (1992b), among others, argue, is not a position of choice which should be adopted at the outset. On the contrary, as Harris (1994a: 187) points out, the null hypothesis is that, wherever syllable structure may be assigned, this process is purely structure building and never structure changing. This is the position adopted by Giegerich (1992), for example, and, due to the existence of the Projection Principle (13), also in all GP work. It seems that this approach is becoming more wide-spread, as it also characterises recent developments such as phonological versions of unification grammars (see e.g. Coleman 1992 or Walther 1993 for discussion), where rules are monotonie, that is, they may add information to representations, but they are unable to remove anything. For the purposes of my investigation of FOD in this and the remaining chapters I shall assume that phonological events are a direct consequence of the building of licensing relations. These relations are built exhaustively as soon as a morphological domain becomes available. What has to be borne in mind is that the Projection Principle (see 3.2.2) precludes any undoing of already established licensing relations.

99

3.2.2 'Ground rules' and some principles of grammar Before concluding this veiy short general introduction to GP, I would like to present and discuss a few 'ground rules' to which work within the theory is meant to adhere, as well as some principles of grammar. The discussion of the 'ground rules', with which I shall begin, follows KLV 1990 veiy closely. The term 'ground rules' is used by KLV (1990: 194) to refer to what may equally well be described as 'metalheoretical principles', at least, that is how I interpret them. For the sake of brevity, I will simply refer to them as principles. They are privativeness, (phonetic) universality and non-arbitrariness (see KLV 1990: 194). Privativeness is used in the traditional (Trubetzkoyan) sense, meaning that all oppositions contrast the absence of a particular property (or element) with its presence. Negative values, such as those familiar from binary features, do not exist. The absence of a particular property means precisely that, i.e. that the property is not present. As a result, what would be a negative value in a binary feature framework is predicted to be inert in GP, since active behaviour (e.g. spreading) by something that is not there is impossible. Analyses of phonological events which crucially make reference to both values of a single feature (plus and minus) constitute apparent counterexamples to this approach. However, as has been shown by, for example, Vago (1988), van der Hulst (1988) and Harris (1994b), such accounts are frequently amenable to reanalysis with reference to only one value of the same prime. Whether all accounts which are problematic for the GP approach can be dealt with in this way is an empirical question which is impossible to answer at present, although there is no principled reason why this should turn out to be impossible. Universality has two manifestations, one mainly phonetic and the other phonological. Firstly, it requires that the same phonological object be manifested uniformly, regardless of the phonological system in which it occurs. In other words, a phonological object is recognisable and interpretable without reference to other members of a given system. This also means that an object which is defined by the phonology as, say, [u] will have a roughly similar acoustic pattern in any language in which this object can be found. It has to be borne in mind here, though, that universality does not involve straightforward articulatory equivalence in all cases. The object which in Japanese, for example, functions precisely like [u] from a phonological point of view (by spreading onto [h] to create a voiceless bilabial fricative or by combining with the element A to yield [o]; see 3.3 for details) is in fact an unrounded high back vowel (Shohei Yoshida (p.c.)). Harris & Lindsey (1995) observe that some aspects of phonetic variability between languages are not phonologically significant, while others appear to involve phonological categories. The question is where exactly we need to draw the line. Specifically, how much phonetic variation is to be permitted for one object to be treated as phonologically the same as another? The sort of issues to be addressed here would be, for example, the question of the 'identity' of English [i:] with [ii], which are essentially regional/social variants of one another. My view is that the phonology should not distinguish between these two, but what exactly are the sort of distinctions that the phonology should reflect?

100 This question cannot be answered satisfactorily at the moment, but as our understanding of the acoustic as well as purely phonological properties of GP elements grows we may at least be travelling towards an answer.13 The second aspect of the principle of universality is that markedness is universal. There is no way in which one could claim a particular phonological entity to be marked in one system but not in another. The third metatheoretical principle in GP is that of non-arbitrariness. It stipulates that there must be 'a direct relation between a phonological process and the context in which it occurs' (KLV 1990: 194). As pointed out by KLV (ibid.), this is a principle to which essentially any full-blown autosegmental framework can adhere by virtue of the sort of representations that are used. This also holds for GP, with one possible exception. This exception is what is known as 'ambient elements' (see KLV 1990: endnote 16; Charette 1991: 75). Ambient elements are phonological primes which manifest themselves by combining with an element or elements present in a given lexical representation, but which have no identifiable local source. Charette (ibid.), for example, argues that it is impossible for a nuclear position containing only the neutral element (see 3.3.1) to be realised in French. In order for such a position to be able to receive phonetic interpretation, the neutral element has to combine with the element A. In the majority of cases, however, there is no local source for A available, so that the existence of an ambient A has to be assumed for French. In spite of the fact that ambient elements, which can essentially be posited at will, are in conflict with the principle of non-arbitrariness, it seems that we have to live with them for the time being, until we gain a better understanding of the true nature of the events involving such elements. It is, however, worth noting that the arbitrariness inherent in positing ambient elements is not unique to GP. On the contrary, it can also be found in other theories. Any vowel epenthesis rule, for example, is arbitrary in precisely the same way. The point is that in other frameworks this is seen as unproblematic. After these three 'ground rules' I will now introduce two principles of grammar, viz. the Projection Principle and the Obligatory Contour Principle. As already mentioned, the Projection Principle (13) ensures that no licensing relations (and thus syllable structure) are destroyed during the progression from underlying representation to the final output of a derivation. According to this principle, the phonology cannot manipulate licensing relations. (13)

13

Projection Principle (KLV 1990: 221 ) Licensing relations are defined at the level of lexical representation and remain constant throughout a phonological derivation.

Recently, weaker versions of what could be seen as another aspect of the principle of universality are being adopted (e.g. in Charette & Göksel 1994), where what appears to be the same segment phonetically is interpreted as two different phonological entities. Clearly, this is not a view one would wish to take without strong evidence, but it seems that more and more evidence in favour of it is forthcoming. To some extent, the proposals to be made in the present work run along the same lines.

101

The fact that the Projection Principle precludes any changes in licensing relations during the course of the derivation means that resyllabification is impossible. This is desirable, because it makes the framework much more constrained than any theory which countenances resyllabification. Note that the Projection Principle allows for licensing relations to be added to in the course of the derivation, while changing existing licensing relations is prohibited. This interpretation of (13) is required for handling analytic morphology (see below), i.e. analytic affixation and compounding, and stress assignment. As far as analytic morphology is concerned, on the first cycle, licensing relations hold within a domain only. On the second cycle, however, additional skeletal points become available, and there is evidence to suggest that new licensing relations are established which involve skeletal positions formerly separated by a domain boundary. Such evidence comes from stress assignment, for example, which entails building licensing relations at various levels of nuclear projection. It stands to reason that this can only occur once word formation has been completed. Let me briefly outline the different kinds of morphological structure which play a role in GP (see Kaye 1993 for a more detailed discussion). Morphological domains (also known as analytic domains) would typically be delimited by #(word-)boundaries in an orthodox SPE-type framework. A morphological domain may contain a +(morpheme- or formative-)boundary, but the phonology is not sensitive to +-boundaries. In other words, +-boundaries are treated by the phonology as if they did not exist. This is, of course, a clear difference between the ways in which this boundary is used in GP and in the SPE (pp. 364ff.) framework itself. In the latter, the +-boundary was often referred to in phonological rules. For example, in GP, a denominal adjective such as #parent+al# is interpreted as a single morphological domain by the phonology. Morphology involving no boundaries other than +-boundaries is, therefore, known as non-analytic morphology. There are essentially the following three morphological configurations which play a role in GP, most of which can, of course, be further expanded by concatenation. The first configuration consists of a single analytic domain, e.g. [A] or [A+B], (Analytic boundaries appear as single brackets [ ] and non-analytic boundaries as +.) An example of [A] would be a morphologically simple word such as [boy] or [go], [A+B], by contrast, shows the morphological structure of a word such as the above-mentioned [parent+al]. Non-analytic affixation in GP corresponds very roughly to level 1 morphology in the framework of Lexical Phonology (see 2.4.3.2 for references). A suffix such as adjectival -al, for example, would be treated as non-analytic in GP because its presence affects the location of primary stress, with primary stress falling on the initial syllable in parent but on the penultimate syllable in parental. Apart from affecting stress assignment, non-analytic affixes are typically unproductive and exhibit a certain degree of lexical selectivity. They may also be associated with phenomena such as so-called closed-syllable shortening (e.g. keep vs. kept in English; see Kaye 1993 or Brockhaus

102

1995 for discussion). Analytic affixes, on the other hand, are usually stress-neutral,14 productive and exhibit no lexical selectivity. No closed-syllable shortening effects are observed in the context of analytic affixation. Analytic affixation characterises both the second and third morphological configurations provided for in GP. The second involves cases such as [[A]B],15 which contains two analytic domains. This can be exemplified by words such as the regular past tense form [[peepjed]. The root (peep) occupies a domain of its own (A), while the analytic suffix (B) does not, which, among other things, makes the prediction that the suffix is unstressed. Thirdly, there is the three-domain [[A][B]]-type configuration, which is best illustrated by compounds, e.g. [[black][board]] or [[tea][spoon]]. Each term of the compound occupies its own domain, so that a full vowel and some degree of stress would be expected for each one. In what follows I will not always distinguish phonological (i.e. licensing domains) and morphological domains from one another explicitly. This is because it is either clear from the context what is meant or because the distinction is unimportant. The second principle of grammar I want to discuss here is the Obligatory Contour Principle (OOP; see e.g. McCarthy 1986 and Yip 1988 for discussion and Goldsmith 1990 ¿p. 307ff.) for an overview of its history). In GP, it has a dual role to play. On the one hand, it ensures that no identical segments appear adjacently on the melody tier and, on the other, it performs a function which differs rather from this original idea of what the OCP should do. This function, which appears to be unique to GP, is to eliminate a nuclear position which occurs next to another nucleus. Some languages, such as English and German for example, restrict this event to configurations involving parametrically licensed domain-final empty nuclear positions, while others, such as French and Yoruba (see Harris 1994a: 205) allow this aspect of the OCP to take effect where both nuclei involved have phonetic content. It is not necessary for the segmental content of the two nuclei involved to be identical, although this happens to be the case in the present example. According to Charette (1991: 88ff.), the structure shown in (7b) is derived from that in (7a) by application of the OCP. (For convenience, both are repeated here as (14a) and (14b) respectively.) (14)

14

15

a.

O, I X I l

N! I X I a

Oj

N2 I X I a

0 3 Ks I I X X I I m i

b.

O, I

N2 I x x I I l a

03 I x I m

N3 I x I i

Obviously, this criterion cannot be invoked in languages with fixed stress, such as French and Hungarian. In principle, at least, the structure [A[B]], involving an analytic prefix which does not occupy a domain of its own, is also possible, but see Kaye 1993 and 4.4.2.6 for arguments against this structure.

103

Gussmann & Kaye (1993: 433) provide data from Polish, where the same mechanism appears to apply. In their account, however, the term 'OCP' is not used. Instead, they refer to this event as 'reduction' and define it roughly as in (15). It is important to note that reduction is more restricted than the OCP as just defined. According to (15), reduction would not apply to a form such as that in (14a), so that (14b) would not have been derived. Everything reduction is designed to achieve, by contrast, would be handled by the OCP. (15)

Reduction An empty nucleus and an immediately following positionless onset are removed from any phonological representation in which they occur.

This obviously raises the issue of why Gussmann & Kaye {op. cit.) find it necessary to propose reduction as a separate concept. They observe in an endnote (number 8) that reduction may turn out to "be but an instance of a more general principle. Conceivably, it could be derived from [...] the OCP.' It seems that exactly what the relationship between the OCP and reduction may be is not entirely clear, so that Gussmann & Kaye prefer to assume, at least for the present, that reduction is a separate principle in UG. Clearly, this issue must be investigated further, but this would be beyond the scope of the present book. All we can do is to bear in mind that this is an area within the theory of GP which is in need of clarification. The effects of reduction (or the application of the OCP, if you prefer) are illustrated in (16), adapting an example from Harris 1994a (p. 204), which consists of the English sequence get a. N2 is a parametrically licensed domain-final empty nuclear position and is eliminated, along with 0 3 . (16)

a.

Ο,

N,

02

N2

I I I I x x x I I I g e t

x

03

N3

I I 3 Χ

b.

Ο,

N,

02

N3

I I I I x x x x I εI tI Ia g

The removal of a nuclear position seems to be an operation which affects the licensing relations established earlier and may, therefore, be in conflict with the Projection Principle (13). This issue has been raised in the GP literature on more than one occasion (e.g. in Harris 1994a (p. 204) and Brockhaus 1995), with somewhat tentative conclusions being drawn. The consensus appears to be that the Projection Principle is intended to prevent changes in constituent structure from taking place (see also Kaye 1993 for the same view). Eliminating an empty nuclear position is not in conflict with this, as 'onsets remain onsets and nuclei remain nuclei' (Kaye 1993: 93). Harris (1994a: 204) further elaborates on this basic stance by pointing out that the identity of the nucleus which licenses the onset position 0 2 is immaterial, provided that the licensing requirements of

104

this position are met. This, of course, is still the case, even after the application of reduction/the OCP. Having provided a general introduction to some of the central ideas and principles of GP, I would now like to return to the main concern of this chapter, i.e. the question of what exactly happens to an obstruent when it undergoes FOD. Obviously, the answer can no longer be [+ voice] -»• [- voice], as it was in the majority of SPE-derivative analyses presented in chapter 2, since binary features are not part of GP. Before making any proposals about what FOD may mean in GP terms, I will first take a look at what exactly is available in GP, that is, what the phonological primes are, what particular properties they exhibit and how they form segmental representations.

3.3 Phonological elements and segmental representations in GP 3.3.1 Phonological elements in GP The phonological primitives in GP are cognitive entities which have both acoustic and articulatory aspects. As already mentioned, they are known as elements, and all segments are either elements themselves or consist of a combination of elements.16 A combination of elements is known as an 'expression'. Elements are unary primes which enjoy realisational autonomy, that is, they are pronounceable in isolation. A by-product of this property is that phonological representations are phonetically interpretable at all levels of derivation. Nothing is filled in, in order to convert a phonological representation into a phonetic one. There is no difference in kind between a lexical representation and the output of a derivation. Each is, in principle, phonetically interpretable. This is captured in the Uniformity Condition (Kaye 1993: 92) in (17), which rules out any kind of underspecification in GP. (17)

Uniformity Condition Phonological representations are directly interpretable at every level.

The most appropriate number of elements is still a matter for debate, but recent work (e.g. Harris & Lindsey 1995) suggests that no more than ten are needed. These are shown, together with their independent realisations in (18), which has been slightly adapted from Ingleby et al. 1994a.

16

See especially KLV 1985, KLV 1990, Harris 1990, Kaye 1990a and Harris & Lindsey 1995 for further details.

105 (18) Element

Articulatory Salient property

Unmarked properties

Acoustic Label

Manifestation

U

Μ

labial

back, high, tense...

Falling

Fall in spectral amplitude

R

[r]

coronal

tap,...

Rising

Rise in spectral amplitude

I

[i]

palatal

non-labial, high, tense...

Extremity

Large spectral gap

A

[α] non-high

non-labial, tense,...

Mass

Convergence of F, & F2

@ M

none

non-labial, back, lax...

Centrality

Central spectral activity

h

[h]

narrowed

glottal,...

Noise

Aperiodic spectral change

?

[?]

occluded

glottal,...

Edge

Abrupt spectral change

Ν

[q]

nasal

non-labial, back...

Murmur

Low frequency intensity

L

L

slack vocal folds

Voicing

Fall in pitch (F0)

Η

H

stiff vocal folds

Voiceless

Rise in pitch (F0)

An expression, arrived at by the compounding (or 'fusing') of elements, has a head and one or more operators. The operator elements contribute only their salient properties, while the head element contributes all those of its properties which are not overridden by the corresponding salient properties of operators. The only element which does not have a salient property is what Harris (1994a: 103113, passim) refers to as the neutral element @.17 The implication of this is that fusion with the neutral element in the operator role results in no change to the head at all. The presence of the neutral element only manifests itself when it itself is the head. The defining properties of the elements shown in (18) are couched in both artìculatory and acoustic terms. The former are provided mainly for the convenience of readers more familiar with articulatory categories. The theoretical commitment of GP, however, is to acoustic properties of elements (see e.g. Lindsey & Harris 1990 and Harris & Lindsey 1995). This reflects the fact that sounds are the experience speaker and hearer share, that perception is prior to production in language acquisition and that it is with reference to cognitive categories based on sound patterns that the child makes sense of what she hears. Readers who are interested in further details of the acoustic definitions of GP elements (and the way in which the appropriateness of these definitions can be tested through their use in automatic speech recognition) are urged to consult the rapidly growing literature 17

In the older GP literature (e.g. in KLV 1985, Charette 1991 and Harris 1990), this was known as the 'cold vowel' v°. Note that the account of @ given here is somewhat simplified. See 4.4.2.6 for more detailed discussion of the status of this element.

106

in this area, especially Lindsey & Harris 1990, Williams & Brockhaus 1992, Harris & Lindsey 1995 and Ingleby et al. 1994a, b. As most readers are likely to be more used to the articulatory definitions of phonological primes which have dominated phonological theory since the time of SPE I shall review these very briefly. The discussion here follows closely proposals made in the GP literature, especially in Harris 1990 and Harris & Lindsey 1995. The first five elements listed in (18), viz. U, R, I, A and can be understood as place or resonance elements. In the context of consonants (i.e. segments which occupy non-nuclear positions), U is associated with labials, e.g. bilabial and labio-dental stops, affricates, fricatives and approximants, R with coronals and I with palatals. The neutral element when playing the role of the head in the fusion operation, specifies velarity in consonants. Which place of articulation the element A should be taken to represent is less clear. A number of proposals have been made for this both in the literature and less formally.18 In my view, A contributes backness and possibly some degree of lowness to consonants, so uvular and pharyngeal segments are likely to contain it. The next three elements shown in (18), viz. h, ? and N, could be classified as manner elements. The 'noise' element h is present in the representation of segments which are associated with a narrowing in the vocal tract which is sufficient to produce noise. It is thus present in fricatives and affricates as well as plosives which exhibit a release burst. Unreleased plosives, such as the velar plosive in the German word Akt ([?akt], 'act'), by contrast, lack h. The 'occlusion' or 'edge' element ? is associated with medial noncontinuants, that is, nasal and oral stops as well as laterals. Finally, the 'nasal' element Ν refers to a lowering of the velum and is found both in nasal stops and nasalised vowels. The two source elements L and H, which relate to laryngeal activity, can form part of segments occupying nuclear and non-nuclear positions. The former are traditionally interpreted as low toned (L) or high-toned (H) vowels/resonants. That is, L denotes slack vocal folds and Η stiff vocal folds.19 For non-nuclear positions, the presence of L is often treated as voicing, while Η indicates voicelessness. The absence of either L or Η is generally interpreted as the absence of an active laryngeal gesture (as in a neutral segment, e.g. a lax stop in Korean).20 As we shall see later in this chapter, these definitions need to be further refined. The more elements are present in the representation of a segment, or, more accurately, the greater the degree of fusion contained in a position (Harris 1994a: 131), the more complex the segment is considered to be. Complexity, in turn, can be roughly equated

18

19

Broadbent (1991), for example, has suggested that A is contained in the representation of English M (realised as [J]), while Brockhaus (1994), disagreeing with this view, argues for treating A as part of consonantal III in languages such as French and German (i.e. [R], [K] or [if]). Jonathan Kaye (p.c.), on the other hand, has proposed that A should be seen as contributing to the composition of pharyngeal/phaiyngealised consonants in Arabic, for example. See Halle & Stevens 1971 for a discussion of these terms. See, for example, Hirose et al. 1974 and Kagaya 1974 for experimental studies which show the absence of an active laryngeal gesture in lax Korean stops.

107

with the traditional notion of phonological strength,21 which is the inverse of sonority. Segments with a higher degree of complexity are generally more likely to act as governors and, according to the Complexity Condition (19), a governee may not be more complex than its governor. ( 19)

Complexity Condition (Harris 1990: 274) Let α and β be segments occupying the positions A and Β respectively. Then, if A governs Β, β must be no more complex than a.

3.3.2 Segmental representations of some German obstruents Elements are arranged on autosegmental lines in such a way that each element occupies its own line (labelled according to the salient property of the element).22 How the autosegmental lines are arranged with regard to each other is currently a matter for debate among GP researchers. The approach with the longest tradition (first proposed in KLV 1985) is that phonological representations consist of a two-dimensional grid where autosegmental lines and segmental positions intersect. This view provides no way of capturing the fact that certain groups of elements can be shown to be consistently targeted by phonological processes. For example, as argued by Harris (1994a) and (in much less detail) by Harris & Lindsey (1995), place assimilation, which can be observed in nasal + stop clusters, say, in English and German, involves one or more of the resonance elements U, R, I, A and and it would therefore be advantageous to be able to refer to them by means of some class label or single node in a tree structure of the type used in feature geometry (e.g. Clements 1985, McCarthy 1988). This is particularly important it as Harris (1994a: 127) argues, phonological processes are to be constrained by the principle in (20). (20)

21

22

23

24

Each phonological process can access only one unit in a representation.24

See 2.3 for references concerning sonority hierarchies and phonological strength scales. Clements 1990 contains a particularly useful recent survey and some interesting proposals on sonority. Lines may be conflated to prevent the elements occupying these lines from fusing with one another and thus to account for the absence of certain segment types, e.g. of front rounded vowels in (conservative) RP (conflation of U- and I-lines). This is but one mechanism for reducing the range of possible segments in a particular language. In recent work (e.g. Charette & Göksel 1994), a more sophisticated means of achieving similar effects is being developed, namely the so-called licensing constraints. Harris (1994a: 126ff.) proposes to interpret palato-alveolar segments as being characterised by the presence of both R and I. This view is in conflict with proposals in Harris & Lindsey 1995, where only I is deemed to be part of these representations. It is the latter approach which will be adopted here. This principle reflects the spirit of proposals made in Clements 1985, Piggott 1988 and Clements & Hume 1993.

108

Although the introduction of geometric structure into GP entails complicating the theory, it seems to me that proposals to this effect are to be welcomed, since the evidence adduced in favour of this strategy (mainly on the basis of binary features, but still partly transferable to GP-type primes) is considerable (see e.g. McCarthy 1988 for a survey and Kenstowicz 1994: chapters 4 and 9, for a more up-to-date discussion). Harris (1994a: 129) proposes a geometric arrangement along the lines of that shown in (21). (21)

χ

This relatively simple tree provides an easy way of formalising a range of phonological processes, including place assimilation (expressible as sharing of the RESONANCE node by adjacent segments), debuccalisation (loss of the RESONANCE node), loss of voicing contrast (loss of the LARYNGEAL node) and loss of segmental content (delinking of the ROOT node). However, the principle in ( 2 0 ) , which stipulates that only one node or element may be accessed by a single phonological process, appears to be too restrictive, as counterexamples which do not readily lend themselves to reanalysis are easy to find. Some of these are contained in recent work by Harris himself (e.g. Harris 1990, 1994a: 212ff.). One involves the vocalisation to [j] of a stem-final underlying palatal stop in Arbore (eastern Cushitic; Ethiopia). Under Harris's analysis, this process involves the loss of two elements, h and ?. As a glance at (21) will confirm, segmental reduction, then, affects two elements. Although both these elements are immediately dominated by the ROOT node, it is not possible for them to be accessed via this node. They have to be delinked individually, because delinking the ROOT node to capture both of them with reference to a single unit would, incorrectly, result in segmental deletion rather than vocalisation. Harris's (1990,1994a: 212ff.) account of f-glottalling in English (e.g. in the dialect spoken in London) is equally incompatible with (20). It involves the delinking or deletion of both R and ?. These two elements are immediately dominated by two separate nodes (RESONANCE and ROOT respectively), so there is no way in which both could be referred to as a single unit in the representation.

109 In principle, there are three basic strategies for dealing with this problem. Firstly, one could abandon (20) altogether. However, this would have the disadvantage of making it impossible to account for the fact that phonological processes typically access only one geometric unit. In other words, we would lose an important reason for setting up a geometric arrangement of elements in the first place. Secondly, it may be possible to make appropriate changes to the geometric arrangement. For example, one could perhaps overcome the difficulties posed by the Arbore vocalisation case through the introduction of an additional node, say, MANNER (as assumed in Clements 1985 and rejected in McCarthy 1988). On the other hand, the arguments against such a node put forward by McCarthy (1988) suggest that this would not be the most fruitful avenue to pursue. The third option would involve a reinvestigation of processes which constitute apparent counterexamples to (20). For instance, what looks like a single process (vocalisation) in Arbore may prove to consist of more than one stage, in which case only one unit per individual process would be accessed. This solution is certainly available for f-lenition in English, which Harris (1990: 292f.) shows to involve two independently motivated stages, viz. loss of release (h) and loss of oral gesture (R). To the extent that such alternative analyses can be developed in all relevant cases, the third strategy seems the best, as it enables us to keep both the relatively simple tree in (21) and the constraining principle in (20). In the remainder of the book, I adopt Harris's geometric approach and assume that (20) is in force. I also follow him (1994a: 101) in occasionally compressing fully articulated representations into two-dimensional grid representations which do not include the three nodes shown in (21). This is done for expository convenience only, and more detailed representations will be used where appropriate. Given the discussion so far, we can set up the following (compressed) segmental representations for the obstruents which are associated with FOD in German. For reasons of exposition I have chosen only a single series, the voiceless one. This, of course, is not meant to prejudge the issue of what FOD is. The palato-alveolar fricative25 is included only for the purpose of giving a comprehensive account, but, in the light of the arguments put forward in 1.2, will not feature any further in this book. Throughout, I shall adopt the practice of identifying heads of expressions (by means of bold type) only where this is immediately relevant to the discussion.

23

The reader may wonder how the voiceless palatal fricative [ç] is to be represented if the palatal element I is already reserved for [J], This can be dealt with in terms of headship. The palato-alveolar fricative has h as its head, which would account for this segment's stridency. The non-strident palatal fricative [ç], by contrast, has I in the role of the head. Apart from this, the two segments do not differ in their phonological representations.

110 (22)

χ

I

υ

U @ ?

?

I I

?

h

h

h

h

H

I I

H

H

H

[Pi

M

[k]

I I

I

[f]

[s]

[J]

As already mentioned, the representations in (22) are those of the voiceless series containing H and characterised by stiff vocal folds. According to KLV 1990, the voiced series would have L rather than H (representing slack vocal folds), and the neutral series, with no active laryngeal gesture, would lack both L and H. There is something of a problem with the representations in (22), concerning the element R. This element is at times referred to as the 'coronal element', since its independent realisation is generally taken to be a coronal tap (e.g. in Lindsey & Harris 1990). It seems obvious, then, that it should be present in the representations of coronal segments. However, it is a well-known fact that coronals have a number of special properties which set them apart from other segments. These special properties are discussed in detail in Paradis & Prunet 1991 and are conveniently summarised by Harris (1994c) as follows: (23)

Coronal syndrome Cross-linguistic incidence (a) Coronals constitute the most frequent class of consonant in the world's languages. (b) All languages have at least one coronal consonant. (c) Liquids are overwhelmingly coronal. (d) The coronal class shows a greater richness of place and manner contrasts than other consonant classes. Processes (e) Consonant harmony occurs exclusively on coronals. (f) Coronals, unlike other classes, are never opaque to harmony. (g) Coronals are more prone to assimilation than other classes.

Ill Phonological distribution (h) (i) (j) (k)

Coronals (typically/universally?) cannot appear independently in the first C of an internal -C.C- cluster. In many languages, a word-final consonant can only be a coronal. In many languages, the margins of word-edge clusters can only be occupied by coronals. Coronals are favourite 'fillers' in hiatus-breaking onsets.

Given this extensive list of properties unique to coronals, it seems wholly inappropriate to invoke a coronal element R and so effectively accord coronal segments the same status as, say, labials (which are characterised by the element U). In line with general thinking in phonology, the need for a way of capturing what is special about coronals has been felt by GP researchers for some time and concrete proposals about eliminating R were first made by Backley (1993), soon to be followed by independently developed ideas from Szigetvári (1994) and sophisticated improvements on Backley's suggestions by Harris (1994c). All three agree that the element R should be dispensed with, but they have differing views on how to deal with the problem of finding alternative representations for coronal segments in the absence of R. Having discussed Backley's proposals elsewhere (in Brockhaus 1994) and found them inadequate for dealing with the consonant inventory of German, I have decided to adopt Harris's (1994c) solution to the problem. Using a geometric arrangement very similar to that in (21) as his point of departure and adopting an approach akin to that of, for example, Avery & Rice (1989), he argues that coronal segments should be treated as lacking a place or resonance element altogether, but as still maintaining the RESONANCE node. That is, the representation of the voiceless coronal plosive would take the form in (24) rather than that in (22).

[t] According to Harris (1994c), the RESONANCE node is like an element in that it defines a privative opposition and enjoys realisational autonomy (i.e. is pronounceable on its own). Unlike elements, however, it is a class node and phonologically inert. The elimination of R, of course, also affects [s], which must now be represented as shown in (25).

112

(25)

χ

h

7

LARYNGEAL

H

RESONANCE

·

[s]

3.3.3 Segmental representations of some German sonorants Having considered the representations of the obstruents relevant to the discussion of FOD, we can now turn to those segments which can precede alternating obstruents in environments where FOD applies. Recall that FOD effects can typically be observed after a (usually long) vowel or diphthong or after orthographic /, η or r. The purpose of investigating the composition of these segments is the following. Where the alternating obstruent is preceded by /, η or r, it can be assumed that the obstruent itself occupies an onset position, while the preceding segment is associated with a rhymal adjunct (or coda) position. Asfirstshown in (3), and illustrated here by the representation of halb ([halp], 'half) in (26), an onset governs a preceding coda. (The broad transcription reflects the actual realisation rather than the underlying form - a practice I shall maintain throughout this book.) (26)

R O X

X

Ο

Ν

X «- X

X

If it turns out that the application of FOD causes a change in the segmental representation of the affected obstruent, then this change will have some effect on this obstruent's ability to discharge its governing responsibilities. In order to be able to judge whether the codaonset governing relation is in any way compromised by the application of FOD, we need to consider the segmental representations associated with both the governing onset and the coda position, to which I shall return presently.

113

Where the alternating obstruent is preceded by a single monophthong or heavy diphthong ([ai], [au] or [oi]), by contrast, no governing relation exists between the onset position occupied by the alternating obstruent and the preceding segment. This is illustrated in (27), where the representations of Dieb ([di:p], 'thief) and Haus ([haus], "house') are shown. Under these circumstances, the composition of the segment preceding the alternating obstruent is immaterial. (27) a.

Ο

Ν

I χ I d

/ \ χ χ \ / i

O N I χ I ρ

b.

I χ

O

Ν

I χ

/ \ χ χ

O N I χ

ι χ

It is, therefore, necessary to focus on the representations of those segments which are associated with coda positions preceding onsets occupied by an FOD obstruent. My initial suggestions for these, as they occur in words such as halb (see (26)), Bund ([bunt], 'league') and Korb ([koçp], Tjasket'), are set out in (28). (28) a.

c.

ROOT

ROOT

ROOT

Ν

RES

RES

RES

Ά

( A ))

[1]

\

[n]

Given the assumptions about the geometric arrangement of elements and the absence of a coronal element just spelt out, the representation of [1] in (28) seems reasonably straightforward, while the representations of [n] and [ç] merit further discussion. Let me begin with the alveolar nasal [η], I have chosen not to include the laryngeal element L in the representation of this segment, although there appears to be evidence which suggests that L is in fact present in all nasals. Such evidence comes from languages where nasals spread voicing onto adjacent obstruents, or at least onto adjacent

114 stops. Spreading of voicing is attested in languages as diverse as Japanese (Itô & Mester 1986), Modern Greek (Pagoni 1993), Kikuyu (Niger-Congo; northern, eastern and central Africa; see Clements 1985), Javanese (Malayo-Polynesian; eastern and central Java; see Hayward 1993) and Zoque (Penutian; Mexico; see Goldsmith 1990: 221). Given the representation of a nasal proposed in (28), this kind of spreading remains entirely mysterious. There is no non-arbitrary way in which it can occur. If, however, an L were included in the representation of the nasal, it would be immediately apparent how voicing could be spread from such a nasal to an oral stop.26 One could claim that it is the laryngeal element L which spreads. There are two options for accounting for the absence of such spreading in languages like English and German. Either the L cannot spread or it is absent from the nasal in the first place. Assuming that the spread of voicing is indeed a consequence of the spreading of L, the former approach would be preferable, at least on the face of it. Firstly, Aere can be no doubt that a great deal of cross-language variation exists with regard to which elements may spread and under what circumstances. So, a mechanism for dealing with this is required in any case.27 Secondly, the principle of universality (see 3.2.2) would require us to assume identical representations for what appear to be acoustically identical objects in different languages, namely voiced nasal stops. However, adding L to the representation of a nasal would raise its complexity to the same level as that of a plosive. In fact, the bilabial nasal [m] would be more complex than the coronal plosive [t] and thus phonologically stronger. There are no sonority hierarchies/segmental strength scales (see 2.3. for references) which rank nasals higher than plosives. To the extent that such hierarchies or scales capture valid insights which should be reflected by segmental representations in GP, implying that nasals may be stronger than plosives would be undesirable. It is, perhaps, the case that it is not L which is responsible for the spreading of voicing from nasals after all. Current research work at the School of Oriental and African Studies (University of London), which is as yet unpublished, suggests that there may be an intimate link between L and N. In a sense, the two may be one and the same element, in which case the representation proposed for [n] in (28b) could be maintained and the absence of the spreading of voicing from nasals in German be handled by an appropriate parameter. This is the strategy I shall adopt in the remainder of this book. Let me now turn to the vocalic /--reflex [ç] and attempt to justify my claim that it is composed of the elements @ and A. First of all, I need to explain that A is bracketed in (28c) to indicate that it is unlicensed. The details of the licensing of individual elements will be discussed in 4.3, so suffice it to say at this stage that an unlicensed element can be thought of as having the same status as an element which is delinked, i.e. not attached to a skeletal position and, therefore, uninterpretable. At the same time, though, an unlicensed element remains available for interpretation should the relevant licensing conditions change for some reason. Whether A is licensed or not determines whether we 26 27

To my knowledge, this approach was first proposed by Jonathan Kaye (p.c.). Piggott (1988) proposes a parameter which is designed to provide precisely such a mechanism.

115

hear consonantal [r] or its vocalic reflex [ç]. (Recall that [r] does not have its IP A value but stands for [Κ], [Ç], [R] or [r], all of which do have their usual IPA values.) In other words, the segmental representation of consonantal [r] is that shown in (29). (29) ROOT

RES

As discussed in 1.2.2.1, orthographic r is realised as a vowel ([ç]) in my dialect, unless it immediately precedes a filled nuclear position. No other consonant can be vocalised in this dialect. (The same is true of the form of German Hochlautung described in the DUDEN pronouncing dictionary (Mangold et al. 1990) and of Northern Standard German.)28 Λ-vocalisation involves alternations between [r] and [ç] in all dialects. This is illustrated by the data set in (30). In (30a), /--final roots without suffixes exemplify the vocalisation of r. The same roots, with vowel-initial suffixes, in (30b) show that the vowel [ç] does indeed alternate with [r].29

28

29

Unschwer bohr' stur fuhr' hör'

M [Jve:ç] [bo:ç] [Jtuiç] [fy:ç] [h0:ç]

'their' "heavy' '(I) drill' 'stubborn' '(I) guide' '(I) hear'

wirr Herr karr' dürr

[viç] [heç] M [dvç]

'confused' 'master' '(I) cart' 'arid'

As noted in 1.2.2.1, Northern Standard German (NSG for short) corresponds to what Vennemann (1972: 12) refers to as 'Standard German, Northern pronunciation'. Recall that a word-final apostrophe in the orthography indicates an apocopated schwa Loss of the 1st sg. pres. ind. suffix -e is extremely common in anything other than careful or emphatic speech.

116

ihre schwere Bohrung sture Führung hörig

[•?i:ra] ['Jveira] [Î>o:ruq] ['/tu ira] ['fy:ruq] [h0:nç]

'their', nom. pl. "heavy', nom. pl. 'drill-hole' 'stubborn', nom. pl. 'guidance' 'enslaved'

wirre Herrin Karre Dürre

['vira] ['her ι η] [Tears] ì'dvra]

'confused', nom. pl. 'mistress' 'cart' 'drought'

My claim, which will be spelt out in some detail in chapter 4, is that particular elements (A in this case) remain unlicensed under certain circumstances. Exactly how the presence of A in [r] can be justified will be discussed in 4.5.3, where I will also address the issue of why a representation which contains only @ as a licensed element is to be phonetically interpreted as [ç] rather than [a]. For the time being, though, I would like to return to the main concern of this book, FOD in German, and begin my investigation of this in the context of GP by considering the changes which it entails.

3.4 What are the changes involved in FOD? 3.4.1 What is 'voicing'? 3.4.1.1 Some general cues to the voicing contrast Asking what the changes involved in FOD are amounts more or less to asking what FOD actually is. Considering its name, final obstruent devoicing, this may look like a completely unnecessary question to pose, but, as I shall show in this and the following subsection, it is far from obvious what exactly the term 'voice' refers to, quite irrespective of any phonological framework. As far as GP is concerned, there is obviously no binary feature [± voice], so I also need to address the question of how voicing phenomena can be expressed in this framework. Before doing so, however, I would first like to take a closer look at the physical phenomena the traditional term 'voicing' is meant to cover. My aim here is to show how complex the acoustic cues to the voicing contrast are, which, phonologically speaking (at least in the traditional terms of [+ voice] vs. [ - voice]), is usually considered to be very simple. The task of identifying invariant acoustic cues to this contrast is enormous and,

117

in spite of a considerable amount of work having been carried out over several decades,30 it is far from completed. So, what are the cues to the voicing contrast, whatever the phonological label one may choose for it? In what follows, I will discuss a number of possible cues, not all of which are suitable for all manners of articulation of an obstruent or all positions within an utterance.311 will always make it clear which cue can be used where. One of thefirstdistinctions between voiced and voiceless obstruents which springs to mind is connected with vocal fold vibration. It would be conceivable to consider only those segments which exhibit glottal pulsing (i.e. vocal fold vibration) throughout their articulation (e.g. throughout the closure phase for stops) as voiced. This, however, is problematic to the extent that some phonologically voiced stops have very little or no vocal fold vibration. Initial and final b, d, g in English words such as bet, debt, get, nib, hid and egg arefrequentlyarticulated with no periodic energy. Furthermore, it would be impossible to explain how hearers are able to identify the voicing value of a whispered obstruent, as whisper involves no glottal pulsing whatsoever. And yet, this task can be carried out with perfect accuracy. Another well-known cue is voice onset time (VOT, originally proposed in Lisker & Abramson 1964). This works best for pretonic (initial) stops, but it can also be used to some extent for intervocalic stops. What is measured is the period of time from the release of the stop to the beginning of periodic energy associated with a following vowel or sonorant. A VOT of around + 20 ms (i.e. the onset of voicing lags behind the release by 20 ms) has been considered to be indicative of a voiceless stop, while 0 or negative VOT values are most typical for voiced stops (at least in a language such as English). In other words, VOT relates aspiration to the voicing contrast. VOT by itself, however, is not entirely adequate because of certain restrictions. Firstly, in initial but not immediately pretonic positions, VOT values are generally reduced and contrasts less clear. Secondly, final stops, whether released or unreleased, cannot be dealt with in terms of VOT, as there is no voicing following these segments. Thirdly, the concept of VOT cannot directly be extended to fricatives, since fricatives are not generally associated with aspiration. There are studies of'VOT' in initial fricatives in existence (e.g. Massaro & Cohen 1977), but what is referred to as 'VOT' here is in fact the period from the onset of friction to the onset of voicing. In other words, the closure duration of the segment itself is included, or, perhaps, the label VOT is even synonymous with 'closure duration' in this context, which marks a clear departure from the original idea. Be that as it may,

30

31

See Watson 1983 for a comprehensive and detailed survey, on which much of the present discussion is based. Further interesting facts and observations can be gleaned from Parker 1974 and Lisker & Abramson 1964 and the references in both papers. By this I mean the traditional tripartite division into utterance-initial, utterance-final and intervocalic position. However, as Watson (1983: 2) points out, this division covers only part of the environments which should be investigated (consonant clusters, word-initial but not utterance-initial position etc. are excluded), but it reflects the bias in the literature. In what follows, the terms 'initial' and 'final' refer to utterance-initial and utterance-final position respectively.

118

there is evidence that shorter duration of friction means that a fricative is more likely to be perceived as voiced (see Watson 1983: 13). Not just temporal cues, such as VOT, have been investigated for initial stops. Spectral properties, such as formant transitions and loci have also attracted attention (see Watson 1983: 4f.), especially with regard to the first formant (F!). Studies reporting links between voicing and spectral information associated with Ft have been challenged in recent years, but conclusive results are not, as yet, available. A further spectral cue concerns the fundamentalfrequency(F0) rather than F,. Some studies suggest that a lowrising F0 contour is perceived as a voiced stop, while others have found additional evidence indicating that it is the onset frequency of F0 which is crucial. Moreover, it has been observed that the spectral information both from F, and from F0 can interact with VOT. Specifically forfinalstops, the relationship between the voicing value of the stop and the length of the preceding vowel has come under close scrutiny. This has been prompted by an apparently universal tendency for vowels preceding a voiced segment to be relatively long, while those preceding a voiceless segment tend to be relatively short.32 In English, this tendency is particularly marked, but there are other languages, such as Spanish and Russian, for example, where it appears that the length difference is too small to be perceptible. So, although vowel length by itself is unlikely to be the sole cue for the voicing value of a segment, it does appear to have some role to play in this context. A further approach to the problem focuses on the transition from the preceding vowel to the following stop.33 The claim being made is that the fundamental frequency of the vowel drops during the transition to a voiced segment (this is termed gradual transition), while it remains unchanged until the onset of the consonant articulation before a voiceless segment (abrupt transition). Again, this may turn out to be a factor in the recognition of the voicing value of a segment, but in itself it is not sufficient, especially as it cannot be used for initial stops, whereas there is some evidence that it does work for final fricatives (see the references in Kohler 1982). As for initial stops, certain properties of the release burst (amplitude, duration and frequency) have been consistently linked with the voicing contrast. Voiceless obstruents tend to be associated with greater amplitude, duration and higher frequency than their voiced congeners. The problem with final stops, though, is that these are often not released, so any cue derived from the release will be unavailable. Moreover, this cue is unsuitable for use with fricatives, as these lack a release burst altogether. Closure duration has been interpreted as yet another cue to the voicing contrast, one of the few cues which are largely unaffected by manner or position (provided, of course, that stops are released). Voiced segments tend to be shorter than their voiceless congeners. What emerges from this discussion is that, as yet, it has been impossible to pinpoint a single acoustic cue or set of cues which consistently and reliably mark a voicing 32 33

See, for example, Chen 1970; also Walsh & Parker 1981a and O'Kane 1978. See Parker 1974, Walsh & Parker 1981b; cf. also O'Kane 1978.

119

contrast for both stops and fricatives, irrespective of the position within the utterance where the segment in question occurs. F0 perturbations were mentioned as one of many potential or partial cues, without much further comment. The reason for this is that I feel that F0 deserves a somewhat more detailed investigation of its own, since it has been explicitly associated with the voicing contrast in GP.

3.4.1.2 F0 perturbations and the voicing contrast In KLV 1990, the claim is made that the laryngeal elements L and H 'control (nonspontaneous) voicing properties in consonants and represent tone on vowels' (p. 216). The unifying property of these apparently disparate aspects of vowels and consonants is F0. It is a reasonably well-attested fact that, in tonogenesis, it is voiced obstruents which give rise to a low tone and voiceless obstruents which can trigger the development of a high tone on the following vowel (see Hombert et al. 1979 and the references there). So, in all likelihood, voiced obstruents are F0 depressers, that is, they are associated with a drop in F0, which manifests itself as a rising contour in voiced consonant + vowel sequences. Conversely, voiceless consonants are F0 raisers. To see the predictions made by GP analyses of English obstruents in the context of F0 perturbations, consider the following segmental representations which have been proposed for English (partly in Harris 1990, partly by Jonathan Kaye (p.c.)).34 I have chosen bilabial stops for the purposes of this example, but the claims made for these segments are also valid for all other English obstruents. To keep things as simple as possible, I have reserved full segmental representations for those segments which are immediately relevant to the discussion and will continue to do so throughout this book. The remaining segments are in broad transcription. Note that the transcriptions of the plosives considered in this section are particularly detailed, so that the different types can be distinguished clearly for the purposes of the discussion. I shall return to less detailed transcriptions at the beginning of the next section (3.4.2).

34

Note that Harris (1994a: 171) now proposes that voiced obstruents occupying head positions in branching onsets do not contain L.

120

(31)

b. blue

a. bee O

Ν

Ν

/\

/ \ χ χ

χ

i

U

V

Μ

Ο

I

[b]

χ I

1

à. pea

e. plea

/\

χ

χ

\ U

/

Ν

0

/\

/\

χ

X

χ

υ

1 1

\ / i

X

0 1 X U

Η

Η

[ph]

[ph]

Ν / \ χ χ \ / i

The claim implicit in this analysis is that those consonants which have traditionally been treated as voiced can be divided into two groups, those which occupy governing positions and those which do not. The [t>] in bee belongs to the latter category, as it has no governing responsibilities and, therefore, may remain less complex (and hence weaker). This is reflected in the fact that it lacks a laryngeal element. It is important to note that the absence o f a laryngeal element does not amount to underspecification, since no filling in is involved at a later stage. The segment is interpretable without either H or L. The b in blue, on the other hand, occupies a governing position in a branching onset. In the unmarked case, the governing position in such a branching constituent is occupied by a segment which is relatively strong, i.e. complex. One way of gaining a higher degree of complexity is to acquire an additional element, the laryngeal element L in the case o f a voiced [b]. Hence the L in blue. Conversely, a voiceless [ph] occupying the governing position in a branching onset is likely to contain H in its segmental representation. This is illustrated in plea. A fourth case (shown in (3 Id)) involves [ph] in a non-branching onset, e.g. in the word pea. Here, H is also present, and marks the lexical distinction between [t>] (as in bee) and [ph] (as in pea). The prediction for F 0 perturbations made by these representations is the following. F 0 should remain relatively constant in words like bee, should show a rising contour in blue and exhibit a falling contour in words such as plea and pea. According to a study carried out by Hombert et al. (1979), this, however, is not the case. In Hombert et al.'s experiment, five adult male American English speakers' pronunciations o f the sentence "Say C[i] again" were recorded, where C = [ph t*1 kh V çl g]. Each token was repeated ten

121

times, and for each one F0 was sampled at 20 ms intervals from vowel onset to 100 ms after vowel onset. The results over all five speakers and all places of articulation showed a falling F0 contour for the voiceless stops (from c. 136 Hz to c. 130 Hz), while a rising F0 was observed for the voiced stops (from 119 Hz to 127 Hz). Considering that the data used in the experiment corresponded to words such as pea and bee, no F0 perturbation would be predicted for bee by the GP analysis set out above, and yet, this is where the (rising) contour was most dramatic. This problem might be solved by positing L as part of what has so far been described as neutral ft)]. On the other hand, this analysis could be in conflict with the GP principle of phonetic universality (see KLV 1990: 194 and discussion in 3.2.2) to the extent that the initial segment in the English word bee differs from that in the French word bis ([bi], 'greyish brown'). Traditionally, the former has been described as partially voiced (or even voiceless unaspirated), while the latter has been classified as fully voiced. Another, more general, problem with treating F0 perturbations as a cue to the voicing contrast is that prevocalic consonants appear to have a very different effect from postvocalic consonants with regard to F0. Hombert et al. (1979: 49), for example, refer to studies which indicate that postvocalic consonants have similar F0 effects on preceding vowels as do prevocalic consonants on following vowels, but with a much smaller magnitude. However, they also mention studies which report a very different result, specifically, that postvocalic consonants invariably lower the F0 in the preceding vowel, regardless of whether these consonants are voiced or voiceless. Reduced magnitude of F0 perturbations may not be a serious problem, as long as the perturbations remain perceptible. Identical effects for both voiced and voiceless consonants, on the other hand, imply that F0 is unsuitable for indicating the voicing value of any word-final (or, at least, prepausai) consonant. That is, unless voiced postvocalic consonants consistently depress F0 to a different degree from voiceless postvocalic consonants, F0 contours cannot be exploited for determining the voicing value of final consonants. This conclusion is further supported by studies referred to in Löfqvist 1975 (p. 237), which report 'no influence of a following consonant on the F0 of a preceding vowel'. These contradictoryfindingsfor F0 as a voicing cue are further complicated by results reported by Parker (1974). He observes that in English words such as pick and pig F035 drops before a voiced stop and remains constant before its voiceless congener. This finding matches that reported by Hombert for prevocalic voiced plosives in English and is problematic for the GP analysis for the same reasons. After all, the GP account predicts arisingF0 contour for pick (since the final [kh] has H) and a relatively constant F0 level for pig (recall that the [g] lacks a laryngeal element). Moreover, Parker reports the same F0 drop before zero, i.e. prepausally, in a vowel-final word. This may indicate that the F0 is not phonologically significant, i.e. that, as far as the phonology is concerned, no such drop occurs at all, which could go some way towards salvaging the GP analysis illustrated in (31). Or, to put it differently, it may show that the baseline is 35

Parker actually refers to the lengthening of the 'period of vocal cord vibration', but, as far as I can ascertain, this is identical with the lowering of the frequency of F0.

122

a slightly falling F0 rather than a steady one. In that case, it raises the issue of exactly how much of an F0 perturbation is required for it to be phonologically significant. On the other hand, Parker's findings may mean that a drop in F0 is not a cue to the voicing contrast at all. Yet another problem is that of the interaction between F0 and stress. Massaro & Cohen (1977: 376) quote a 1972 study by Lea in which F0 perturbations were examined both for stressed and unstressed syllables in English homographs, such as permit, where stress falls on the initial syllable for the noun and on the final syllable for the verb. It was found that the F0 contour alone did not reliably distinguish between voiced and voiceless consonants in these words. A falling F0 contour, for example, was associated both with voiceless and voiced prevocalic obstruents in an unstressed syllable. Clearly, there is substantial room for error in a study of this type. Differing manner of articulation of the consonant in question (plosive vs. nasal in permit, for example) may have distorted the results. Also, it is not clear from the information provided by Massaro & Cohen how Lea avoided including F0 effects caused by the intonation in his investigation of voicingconditioned F0 contours. This, in fact, is the crux of the matter. Given the role F0 plays in stress and intonation, how can it be reliably isolated as a voicing cue? This problem is tackled in Kohler 1982, where it becomes clear that very carefully controlled experimental conditions and relatively complex interpretation algorithms are necessary in order to be able to isolate segmentally conditioned F0 perturbations from intonational macropatterns. Finally, on the basis of an investigation of pitch as a voicing cue, Haggard et al. (1970) hypothesise that the falling F0 contour associated with a prevocalic voiceless stop (which they term 'pitch skip/dip') is likely to characterise also the contrast between [f, v] and [Θ, Ö], but not [s, z] and [J, 3]. The fact that, according to the authors, the glottis opens partially for the latter pairs of fricatives means that cessation of glottal vibration is common and, therefore, pitch skip/dip occurs even for the voiced fricative. So, F0 perturbations may not be a suitable voicing cue for all fricative types. It appears that, although F0 perturbations clearly have some role to play in expressing a voicing contrast, it is far from obvious what exactly that role is and how these perturbations are correlated with the laryngeal elements. It may turn out that it is the interaction of F0 disturbances and VOT, for example, which is the invariant correlate of phonological elements, or perhaps the interaction of F0 perturbations and several of the other cues mentioned in 3.4.1.1. Whatever the outcome may be, there can be little doubt that establishing the precise acoustic correlates of the laryngeal elements requires extensive research work, extensive enough to merit a book of its own. It is for this reason that I shall concentrate on phonological arguments and develop an analysis more or less on the basis of these arguments alone. Given the relatively highly constrained nature of GP, such an analysis should make strong predictions which can be tested at the phonological level, so that phonetic evidence of the kind just discussed may not be crucial. In the light of the fact that the precise nature of'voicing' is still so poorly understood, regardless of whether one attempts to deal with it in terms of acoustic cues or by referring to articulatory facts (such as vocal fold vibration; see the beginning of 3.4.1.1

123

for discussion), I will continue to use the term in its broad traditional sense. More precise terms (such as references to GP laryngeal elements) will be employed where necessary.

3.4.2 What are the options for FOD made available by GP? There are two aspects to the question of how FOD can be interpreted in the framework of GP. Firstly, a decision has to be made about which elements are involved and, secondly, we need to know what kind of phonological process we are dealing with.36 In what follows, I will address both these issues at a fairly superficial level. My aim is to identify die most plausible solution at that level and to subsequently put this solution to the test by pushing it to its logical conclusions and seeing whether it makes the right predictions. Before identifying FOD as a particular phonological process, I first need to establish what phonological processes are available in GP. This framework recognises two types of phonological process, composition and decomposition. Either an element is added to a phonological representation, usually through spreading from a neighbouring segment37 (composition), or an element is lost in some way38 (decomposition, reduction). No other phonological processes are countenanced by the theory. Both composition and decomposition affect the phonological strength of the segment concerned. Phonological strength is a notion which is by no means new and strength hierarchies have been set up on numerous occasions, to account for particular historical sound changes, synchronic clustering asymmetries, typical tautosyllabic segment sequences and so on.39 I will return to other aspects of how exactly GP handles these facts later in this chapter and in the following chapters. What matters for the moment is that something approximating the traditional notion of strength is captured by means of complexity,40 as already observed in the context of the discussion of the Complexity Condition (19). The more elements a segment is composed of, the more complex it is deemed to be. Strength is directly proportional to complexity, so, more complex segments are stronger than less complex segments. This interpretation of strength automatically leads to decomposition processes being equated with weakening, and composition with strengthening. To return to the special case of FOD, it is now clear that treating it as a process means that it has to involve either composition or decomposition. Let me consider the former 36

37 38

39 40

This, of course, begs the question of whether FOD is a phonological process (in the sense that there is an actual change involved) in the first place. It does, however, seem the most obvious point of departure. I will return to this issue in chapter 4. The notion of spreading will be further refined in 3.4.3.2.1 and 3.4.3.2.3. The withdrawal of licensing from individual elements or nodes as discussed in the context of the vocalic r-reflex in 3.3.3 must obviously be included under decomposition. See 2.3 and footnote 21 of the present chapter for references concerning strength and sonority. See Harris 1990 (p. 274) for a more detailed discussion which argues for a slightly different position from the one adopted here.

124

option first. If FOD involves composition, the only possibility is that voicelessness is added to an underlyingly voiced segment. In principle, this could be achieved by adding H. This raises the question of whether the lexical representation of the affected segment would contain L or no laryngeal element at all. In other words, would the addition of H result in an expression containing both H and L or H only? According to Harris (1994a: 135), expressions in which both H and L are present are associated with breathy voice, typical of certain stop series in languages such as Hindi and Gujarati. German lacks stops of this kind, so that it is probably fair to assume that the autosegmental H- and L-lines are conflated in German, preventing the two laryngeal elements from compounding with each other. Such conflation of lines has already been mentioned in the context of the Uand I-lines in RP (see footnote 22.) This forces the conclusion that, under the assumption that devoicing involves composition, a devoiced obstruent contains only H. And this returns us to the original question, that is, the question of whether H replaces an underlyingly present L or whether H is attached to an expression lacking a laryngeal element altogether. Delinking lexical L and attaching H would be an option only if it can be shown that the two operations are independently motivated. Mascaró's (1987a) account of voicing phenomena discussed in 2.2.3 is an example of how reduction-cumspreading can be justified. In his analysis, German exhibits reduction only, while languages such as Dutch and Catalan require both reduction (to account for word-final devoicing) and spreading of [a voice] (to capture voicing assimilation). The original question then breaks down into two more specific ones. What evidence is there for reduction (loss of L) in FOD environments? What evidence is there for spreading of H in the same environments? The second question can be dealt with more simply, so I shall tackle itfirst,before devoting the remainder of the present chapter and most of the next chapter to the first question. To begin our consideration of the evidence for spreading of H in FOD environments, let us, first of all, take a look at some typical examples of such environments, with segmental representations of the segment immediately adjacent to the alternating obstruent and of this obstruent itself. Two such configurations (using Dieb ('thief) and halb ('half)) are shown in (32). The question marks indicate where H would be if spreading took place.

125

(32) a. [di:p]

b. [halp] R

0

Ν

1 /\ χ χ I \/ d I ι

Ο

X

U

Ν

Ο

Ν

χ

χ

h

a

χ

Ο

Ν

χ

χ

U

What the representations in (32) indicate clearly is that there is no local source for Η in words such as Dieb and halb. Given that FOD typically applies to obstruents preceded by a vowel of some sort or a sonorant consonant (but not usually another obstruent), this comes as no surprise, since only obstruents can contain the elements Η and L in their segmental make-up in non-tone languages such as English and German. There can be no question of genuine spreading of Η at all, then. Instead, we may be dealing with the sort of event which was described in 3.2.2 in the context of French schwa. As argued by Charette (1991: 75), this segment contains a so-called ambient A, that is, an element which is not present in lexical representation and for which no local source can be identified, but which nevertheless needs to be present in a particular expression for that expression to receive phonetic interpretation. In the case of French schwa, the segment is postulated to consist of an underlyingly empty nuclear position to which an ambient A is added, yielding [a]. However, positing ambient elements is in conflict with the metatheoretical principle of non-arbitrariness discussed in 3.2.2, so claiming that an ambient Η 'spreads' must remain very much a last resort. Interpreting FOD as a composition event, then, does not appear to be a fruitful avenue to pursue, so that a closer look at the decomposition approach has to be the next step for us. Treating FOD as a case of decomposition would mean that a lexically present element which is responsible for making the affected segment voiced is lost through the application of FOD. In other words, one could say that voiced alternants lexically contain L, and that this element is delinked by FOD. This approach has the distinct advantage over the composition analysis that no ambient element needs to be invoked. Investigating this solution amounts to considering the second question raised earlier, that is, the question of what evidence there is for reduction (loss of L) in FOD environments. As already observed (and discussed in some detail in 2.3), other authors before me have entertained the idea that FOD involves reduction. To some extent, this kind of

126

approach is adopted by Mascaré (1987a), Wiese (1990), Goldsmith (1989) and Lombardi (1991). For Mascará and Wiese, an underlying [+ voice] specification is delinked (but filled in as [- voice] by a default rule or other comparable device later), while for Goldsmith and Lombardi no privative [voice] feature may occupy a coda position in an FOD language. Let me now deal with the question of which elements are involved in a GP account of FOD in more detail. I have already suggested that H and L may have a role to play. This comes as no surprise in the light of the fact that they 'control (non-spontaneous) voicing properties in consonants' (KLV 1990: 216), as mentioned earlier. None of the other elements available in GP are serious candidates for involvement in FOD. I will assume, then, as a working hypothesis, that FOD consists in the loss of L. In the remainder of this chapter, I will explore the implications of this hypothesis in detail and see how well the predictions that can be made on the basis of it are supported.41 Before moving on to this, however, let me make a general point which arises from treating FOD as the loss of L. As observed in 3.3.1, the laryngeal elements H and L are associated with résonants in tone languages, where they represent high tone and low tone respectively. In languages such as English and German, however, they are reserved for obstruents only, where they represent voicing or voicelessness. What flows from these assumptions in the light of the proposed working hypothesis on FOD is that sonorant devoicing should not be attested at all. Either we are dealing with a non-tone language such as English or German, where laryngeal elements are not present in résonants, or we are looking at atone language, where delinking of a laryngeal element would mean loss of tone, not of voicing. This conclusion appears to be incorrect, since sonorant devoicing has been reported in the literature, at least for Angas (Halle & Clements 1983: 45), Russian (e.g. Barry 1989), Icelandic, French (both Escure 1975: 165f.), Danish (Rischel 1991), Turkish (Kaisse 1986), Polish (e.g. Gussmann 1992) and English (e.g. Kiparsky 1985). It is worth noting, however, that sonorant devoicing frequently occurs where the affected segment is adjacent to a voiceless obstruent. In other words, the process may involve spreading of the noise element h42 and thus be quite different from FOD. (A similar view, incidentally, is put forward by Lombardi (1991 : vii).) Where there is no adjacent obstruent present, the 'devoicing' effect does not appear to be phonologically significant,

41

42

My decision to investigate the possibility that FOD involves the loss of L also receives a certain amount of supportfromthe experimental work carried out by, for example, Hombert et al. (1979), who reported a rising F0 contour associated with prevocalic voiced consonants in English. Such a contour may suggest the presence of L. It is worth bearing in mind, however, that the experimental findings are far from consistent and clear cut, which is why the main focus of the present study is on phonological evidence. The idea here is that so-called voiceless sonorants are not voiceless in the sense that obstruents can be, but possess properties of aspirates, which give the impression of voicelessness. What is problematic for my account, though, is the fact that only voiceless obstruents act as a source from which h can spread to a sonorant Voiced obstruents do not appear to have this property - a fact for which I have no explanation at present.

127

as, at least in French, Polish and Turkish, no phonological contrasts are affected. In other words, what has been described as 'sonorant devoicing' in the literature is distinct from FOD in that it either consists in a very different phonological process or has no place in the phonology of the language at all, since it is not linguistically significant. Having made this general point, I will now examine some more specific predictions which my analysis of FOD makes for German.

3.4.3 Final obstruent devoicing as loss of lexical L 3.4.3.1 A typical weakening process? Many phonologists would probably answer the question of whether FOD is a typical weakening process with a definite no. This can be concluded from various strength hierarchies (see footnote 21) and explicit statements43 made in the past. However, as Harris (1990: 257) observes, there is no general consensus on which processes are to be interpreted as weakening and which as strengthening. The same process may be weakening for one investigator and strengthening for another. These inconsistencies are due to the fact that, when examined according to traditional methods, different aspects of phonological strength allow different, even contradictory, conclusions to be drawn. Traditional ways of looking at phonological strength provide a less than coherent picture. In the remainder of this subsection I will discuss a study of phonological weakening which illustrates this point very well. In a 1977 paper (based on her 1975 PhD thesis) Escure presents a survey of widely accepted weakening processes in a variety of languages, e.g. spirantisation, gliding and deletion, on the basis of which she identifies several apparently typical properties of weakening processes. These properties are then expressed in terms of three hierarchies, viz. an environmental hierarchy, which ranks possible environments according to their likelihood to be associated with weakening, a hierarchy of major class and manner features, which captures typical weakening trajectories and thus amounts to a strength hierarchy, and a hierarchy of cavity features, which correlates place of articulation and propensity to undergo weakening processes. The hierarchy which is particularly important here is Escure's environmental hierarchy, reproduced as (33). Environments where weakening is most common/most likely appear at the top and environments where weakening is least likely/unattested at the bottom.

43

E.g. Foley, quoted by Smith (1981: 589) and Lass (1971: 27).

128

(33)

(a) V _ C 1 n # # o r V C 1η" _ # # (b) V __

ROOT

y

RES

RES

ROOl

RES

·

u M

[t]

ra

[s]

The facts of English and German suggest that two separate parameters may be involved. The first determines whether a coda position can license an independent LARYNGEAL node and the second, which is parasitic on the first, enables the coda position to share the LARYNGEAL node of the governing onset (English) or prevents it from doing so (German). Where sharing is permitted, genuine voicing assimilation effects can be observed, whereas codas which do not share the onset's LARYNGEAL node will be

136

exclusively occupied by neutral/voiceless segments of the type shown in (41). The parameters are spelt out in (42). (42)

Coda Parameters Coda Parameter I A coda position licenses a LARYNGEAL node: ON/[OFF] Coda Parameter II

A coda position may share the LARYNGEAL node of its governing onset: ON/OFF The default setting for Coda Parameter I is OFF, which is also the setting we encounter in languages such as English and German. In fact, some authors implicitly (e.g. KLV 1990) or explicitly make the claim that laryngeal elements are universally banned from occupying coda positions or, as Rice & Avery (1990: 430) put it, that 'laryngeal distinctions are universally neutralised syllable-finally'. Whether this is correct is an empirical issue which I am in no position to resolve, so I shall leave open the question of whether Coda Parameter I should be replaced by a universal principle which prevents LARYNGEAL nodes from being licensed in coda positions. Such a solution would be equally compatible with my analysis of FOD in German as maintaining Coda Parameter I. Whatever the best way of dealing with this particular question may turn out to be, Coda Parameter II seems well enough motivated. Evidence of variation between languages along the lines specified by Coda Parameter II is easy enough to pinpoint. As shown in the discussion of the vesper - lesbian and Vesper - Lesbe pairs, English, for example, has this parameter set at ON, while for German and French it is set at OFF. These parameter settings account for a number of distributional restrictions which have in the past been treated as syllable-final devoicing or even voicing assimilation in German (see chapter 2). How exactly this works will be discussed in 3.4.3.2.2. We have now reached the stage at which we can discuss the second problem with Harris's (1994a: 171) implicit claim that all voiced obstruents in English lack laryngeal elements. If that is indeed the case, then we would expect coda obstruents (which contain neither H nor L) to sound voiced, unless they share the H of the governing onset. This prediction can be tested by looking for clusters which involve sonorant onsets governing coda obstruents. (Recall that I assume sonorants to have no LARYNGEAL node, so that there are no laryngeal elements available for the coda obstruent to share.) Such cases are not particularly common, but KLV ( 1990: endnote 21 ) provide the adjective arachnoid as an example. They argue that the [k] must close the second syllable in order to account for the fact that it carries the main stress. In other words, the [k] occupies the coda and the nasal is attached to the governing onset. There is no H in the nasal which could be shared by the plosive and yet the plosive is voiceless. Given the parameter settings in (42) and the claims about constituent structure made by KLV (op. cit.), this form appears to suggest that Harris's interpretation of voiced obstruents as neutral is incorrect. Instead,

137

it lends support to my claim that neutral obstruents are interpreted as voiceless and that, for this reason, the voiced obstruents of English and German must contain L. Let me now return to the implications of the segmental representations shown in (35) and (38). The most obvious of these is that so-called devoiced obstruents are distinct from their lexically voiceless congeners at all levels of representation. Lexically, voiced segments have L, while their voiceless congeners have H. When occurring in an FOD environment, segments with lexical L have no laryngeal element at all. In other words, my analysis of FOD entails the claim that it is not a neutralisation process (in the sense of Kiparsky 1976), insofar as devoiced obstruents are distinguishable from their underlyingly voiceless congeners. This sharply contrasts with the majority of phonological analyses of FOD available to date, certainly of those which employ binary features.50 If my claim is correct, however, then the absence of neutralisation in the sense just defined ought to manifest itself both phonologically and phonetically. The relevant evidence and associated issues will be discussed in more detail in chapter 5, so I will only summarise the most important points here. On the whole, it appears that my claim is borne out by the facts. Speakers of NSG frequently spirantise underlyingly voiced velar plosives in FOD environments, whereas spirantisation of non-alternating (underlyingly voiceless) velar plosives is impossible, as witnessed by the data in (43). This suggests that these speakers maintain a phonological distinction between the two segment types. Hochlautung, unlike NSG, does not exhibit general spirantisation of velars, except for words ending in -ig, as discussed in chapter 1. Spirantisation of underlyingly voiced velar stops in FOD environments, as we encounter it in NSG, is illustrated in (43a). (43b) shows that spirantisation of underlyingly voiceless velar stops is blocked in the same dialect. (43)

a. Lexically voiced (containing L) Berg Tag trug Sog lag Sieg Weg Zeug

[beçç] [tax] [truix] [zo:x] [la:x] [zi:ç] [ve:ç] [tsoiç]

'mountain' 'day' '(he/she/it) carried' 'suction' '(he/she/it) lay' 'victory' 'way' 'stuff

b. Lexically voiceless (containing H) Werk schrak

50

[veçk] [Jra:k]

'factory' '(he/she/it) shrank (back)'

See 5.4.2 for a more detailed discussion of this point.

138

buk Pik

[bu-Je] [piik]

'(he/she/it) baked' 'spades' (cards)

As far as the phonetic evidence is concerned, there are a number of experimental production studies addressing the very point at issue here.51 They report statistically significant (albeit small) differences in the duration of the preceding nucleus, the duration of voicing into the closure/friction phase of the affected obstruent and the duration of that closure/friction phase between lexically voiced and lexically voiceless obstruents in FOD environments. These findings are summarised in (44). (44)

Phonetic variables

Most common findings

Duration of preceding vocalic nucleus Duration of voicing into closure/friction (from the left) Duration of closure/friction

Lexical L H longer shorter longer shorter shorter

longer

This evidence is further strengthened by perception experiments which examined native hearers' ability to distinguish between lexically voiced and lexically voiceless obstruents in FOD environments, without access to context for disambiguation. The findings of these experiments are summarised in (45). (45)

Study

Percentage of correctly identified tokens

Port et al. 1981 O'Dell & Port 1983 Port & O'Dell 1985 Port & Crawford 1989

70% 63% 59% 69%

As already mentioned, a more detailed discussion of these facts and their implications follows in chapter 5.

See Port & O'Dell 1985, Fouiakis & Iverson 1984 (as reanalysed in Port & O'Dell 1985) and Port & Crawford 1989; similar findings are also reported for Polish in Slowiaczek & Dinnsen 1985 and Jassem & Richter 1989, for Russian in Pye 1986 and for Catalan in Dinnsen & Charles-Luce 1984 and Charles-Luce & Dinnsen 1987.

139

3.4.3.2.2 Syllable-final devoicing is not a dynamic phonological process As discussed in chapter 2, a number of distributional restrictions involving voice (or rather its absence) have been dealt with in the context of final devoicing. Rubach (1990), for example, would attribute the voicelessness of the emboldened obstruent in non-native words such as Charybdis and Charisma to the application of a syllable-final devoicing rule. My account, however, makes the application of any dynamic phonological process for this purpose unnecessary, as a glance at the representations of these words will confirm. Recall that, under my analysis, devoiced segments lack the LARYNGEAL node and that, due to appropriate Coda Parameter (42) settings, coda positions may be occupied only by segments which do not have a LARYNGEAL node. In other words, my account makes the claim that devoiced obstruents and obstruents which occupy a coda position are identical in their segmental composition. (46)

a. Charybdis [ ç a t Y p d i s ] R κ 0 1 1

Ν I 1

0 1 1

χ

X

I1 ç

11 a

N\ I \\ 1

0 Ι I

Ν Ι I

0 Ι I

χ

X

χ

χ

χ

χ

I 1

1 1 Y

Ι I d

Ι I

Γ

ι I Ρ

I

Ι I s

0 Ι I

Ν Ι I

b. Charisma [ça"risma] R κ Ο I I

Ν I 1

0 Ι I

χ

X

χ

χ

χ

χ

χ

I1 ç

11 a

Ι I

ι I

Γ

I

Ι I s

Ι I m

Ι I a

NN Ι \\ I

The obstruent in question occupies the coda position and, therefore, must not have a LARYNGEAL node. This requirement (not to contain a LARYNGEAL node) has to be met at the point when licensing relations are established, that is, at the level of lexical representation as defined in 3.2.1. So, the segment is lexically without a LARYNGEAL node and no devoicing process applies to it. This means that lexically there are three obstruent types, viz. those containing L, those containing H and those which lack a LARYNGEAL node altogether. In principle, this could mean that a three-way voicing

140

contrast exists in the lexicon, in spite of the fact that only a two-way contrast can be motivated phonologically (see e.g. Keating 1984). The lexical three-way contrast, however, is just an illusion, as no more than two distinct segment types can occur in any one environment. Obstruents without a LARYNGEAL node are found in two environments. Firstly, they occur precisely where no laryngeal contrast can ever be established at all (that is, in the coda position). Secondly, they are present in true FOD environments, where the only other permissible segments are those containing H. Elsewhere, we find only segments which have a LARYNGEAL node, dominating either H or L. So, no threeway contrast in the traditional sense exists in any one environment. As already observed, treating FOD as the loss of lexical L (yielding an obstruent without a LARYNGEAL node) means that obstruents which are subject to FOD and coda obstruents are identical. The voicelessness of [p] in Charybdis and [s] in Charisma can then be accounted for very simply, as both segments occupy a coda position. Many other examples of so-called syllable-final devoicing conform to the same pattern. The supposedly syllable-final obstruent is in a coda position and, therefore, cannot accommodate a LARYNGEAL node. Given that its segmental representation is identical to that of a devoiced final obstruent, this means that it, too, has to be interpreted as 'voiceless' by the hearer. To extend the data set illustrating distributional constraints regarding voice, the relevant listfromchapter 1 is reproduced here (in a slightly modified form which will be motivated shortly) as (47), for convenience. (47)

a. Native words Ab.laß aus. weichen los.gehen

[?aplas] 'indulgence' [?ausvaiç3n] 'to avoid' ['lo:sge:sn] 'to walk off;

und als

[?unt] [?als]

'and' 'as'

[ça'rypdis] [ça'risma] [?ap'do:man] ['?aplati:f] [iakbi] [•bakdat] ['betmintsn] [maries ismus] [Vikvam]

'Charybdis' 'charisma' 'abdomen' 'ablative' 'Rugby' 'Bagdad' "badminton' 'Marxism' 'wigwam'

b. Fremdwörter Charyb.dis Charis.ma Ab.domen Ab.lativ Rug.by Bag.dad Bad.minton Marxis.mus Wig.wam

141

Simbab.we Wod.ka Us.beke

[zim'bapve] [Votka] [Tuslieika]

'Zimbabwe' 'vodka' 'Uzbek'

Adjektiv

[?atjekti:f]

'adjective'

The pair of conjunctions (und, als) in (a) is of no interest at this point, as FOD applies word-finally52 anyway. What matters is the set of words with internal devoicing, i.e. the remaining items on the list. All the Fremdwörter (in (47b)) can be analysed in exactly the same way as Charybdis and Charisma* Note that not a single one of the obstruents under discussion here is preceded by a long vowel or diphthong. As illustrated in (48a), such nuclei would necessitate the formation of a ternary rhyme if the following obstruent was to occupy the coda position at all. However, as we saw in 3.2.1, the coda in a ternary rhyme typically has no RESONANCE node of its own and shares that of the governing onset. This manifests itself as place assimilation, so the obstruent clusters involved would have to be homorganic. If, by contrast, the vowel is short, that is, the nucleus occupies only one skeletal position (as shown in (48b)), the coda position is doubly licensed (see 4.3) and can then license a RESONANCE node of its own, so that non-homorganic clusters, such as the ones occurring in (47) would be expected. (48) a.

R

b.

R

The first group of native words in (47a), by contrast, is less uniformly amenable to the treatment described for the Fremdwörter. The very first word (Ablaß) poses no difficulties in this respect,54 but ausweichen and losgehen could well be problematic, as the 'devoiced' obstruent could not occupy a coda due to the fact that it is preceded by a heavy diphthong in ausweichen and by a long vowel in losgehen and does not share the place of articulation of the following onset. Interestingly, however, there is every indication that aus and los are analytic prefixes which occupy their own domain. In other

52 53

54

This will be discussed in detail in 4.2.1. There may be a problem with the noun Adjektiv concerning the governing power of the segment associated with the governing onset position. This will be dealt with at the end of this subsection. As far as the constituent structure alone is concerned, one could indeed assume that [pi] in Ablaß constitutes a coda-onset cluster. However, the composition of the two segments involved may well rule this possibility out. The problem is essentially the same as that mentioned in the previous footnote regarding Adjektiv. It will be discussed below.

142

words, we are really dealing with compound forms of the type [[A][B]]. Kaye (1993) argues that prefixation of the *[A[B]] variety is universally impossible, a view which, as we shall see in 4.4.2.6 can be further supported with evidence from German, so the structure suggested here for aus- and los- is not unusual. The fact that they carry the main stress in ausweichen and losgehen indicates that, as far as stress assignment is concerned, they behave precisely like the first term in more 'obvious' compounds, such as Milchmann (['milçman], 'milkman') or Radweg (['raitveik], 'cycle path').55 Furthermore, both aus and los can occur independently, aus as a preposition and as an adjective/adverb, and los as an adjective/adverb only. Examples of this are given in (49). (49)

aus dem Haus aus Langeweile

'out of the house' "because of boredom'

Der Film ist aus.

The film is over.'

Der Hund ist los. Was ist los?

The dog is loose.' 'What's the matter?'

The two verbs, ausweichen and losgehen, are what traditional grammars of German (e.g. Hammond 1981) call 'separable' verbs. This term captures the fact that in the infinitive they appear as a single word, but that the prefix (or rather, the first term of the compound) has to be separated out under certain circumstances, e.g. in present indicative main clauses. How this works is shown in (50). (50)

ausweichen Ich weiche dem Mann aus. Der Fahrer weicht aus.

Ί avoid the man.' The driver swerves.'

losgehen Ich gehe jetzt los. Das Gewehr geht gleich los.

'I'll start walking now.' The gun is about to go off.'

In compounds like this, FOD does actually apply, although obviously not woni-fmally. The mechanism involved will be discussed in 4.2.1, so, at this point, all that needs to be said is that FOD applies domain-imsWy (which would account for losgehen and ausweichen) as well as wwtf-finally. Incidentally, although, in principle, Ablaß could be treated as a case of a coda being occupied by a neutral plosive, the fact of the matter seems to be that we are dealing with another compound, as indicated by the fact that ab- bears primary stress in Ablaß and, in other respects as well, can behave in precisely the same way as aus. In any case, it may 55

See chapter 1 for further examples.

143

well be that [1] is not complex enough to govern [p], so a coda-onset interpretation would be ruled out. As yet, we have not decided how exactly complexity is to be calculated, so I shall return to the [pi] cluster once this has been established - later on in the present section. An alternative to the coda-onset view would be to posit a branching onset. Given the spelling, one would expect such an onset to contain a voiced bilabial plosive rather than a voiceless one. That is, we would predict [bl] instead of [pi], which is, however, not what we find. It seems that the formation of a branching onset is blocked. The morphological structure ofAblaß is, then, as shown in (51), which illustrates the state of affairs on the first cycle. (51)

O

N

O

N

O

N

O

N

I I I I I I I I X X x][x χ χ x]] I I I I I I

[[χ

?

a

ρ

l

a

s

The formation of a branching onset across a domain boundary is blocked universally, as governing relations (which include the formation of constituents) are originally established within a domain. When the second domain, which contains the skeletal point for the potential branching onset, becomes accessible on a later cycle, these governing relations cannot be undone again and, therefore, it is not possible to syllabify two positions from separate domains into a single constituent.56 This would be in conflict with the Projection Principle (13), as discussed in 3.2.2. To sum up so far, I have shown that many cases which have been treated as syllablefinal devoicing in the past actually involve no devoicing at all, but can be seen as instantiations of the settings of two linked Coda Parameters (42), which specify which geometric nodes can be licensed by the coda position.57 Before concluding this section, I would like to discuss a potential problem involving a subset of the words listed in (47), which I hinted at in footnote 53 and which is the reason for Adjektiv having been separated from the remainder of the list. If there is no internal analytic morphology in the noun Adjektiv (similarly, Adjutant 'adjutant', adjustieren 'to calibrate', etc.), then the following governing configuration would hold between the coda of the first syllable and the onset of the second syllable.

56

57

Quite apart from this, there is an empty nuclear position separating the two potential members of the branching onset. This would have to be deleted in some way for the formation of a branching onset to become possible. GP, however, makes no provisions for the deletion of such positions, except under reduction/the OCP, which cannot apply here (see discussion in 3.2.2). As already observed in the context of the discussion of (42), the same facts could be handled by means of a universal principle (capturing the OFF setting of Coda Parameter I) and a parameter (Coda Parameter II).

144

(52)

R

N\

O

Χ

Χ

I

I

\Λ

4-

χ

Ι > ROOT

Τ

h RESONANCE I

[a

t

j]

The question is whether this governing relation between the onset occupied by [j] and the preceding coda, occupied by [t], can actually be established or whether it would be in conflict with the Complexity Condition (19). The answer to this question obviously depends on which parts of the representation count for the purposes of determining the complexity of a segment. There are essentially three possibilities. Either we count only 'genuine' elements, i.e. those which appear in (18), or we count 'genuine' elements as well as class nodes, or we adopt a hybrid approach under which, say, the RESONANCE node is counted only when it is independently manifested in the signal, that is, when it dominates no elements (as in the representation of [t] in (52)). If we opt for the first approach, that is, we do not take class nodes into account with regard to complexity, the result would be that, from a complexity point of view, the segmental representations in (53a) and (53b), both of which represent a coronal tap [r], would be equivalent. Or, to put it differently, it does not matter whether the LARYNGEAL node is present or not. (53)

a.

χ

b.

χ

ROOT F LAR RES

·

RES

·

[r] However, as we have seen in the context of the discussion of the Coda Parameters (42), languages do care about the presence or otherwise of a node. German and English codas,

145

for example, do not tolerate the presence of a LARYNGEAL node58 (as captured by the setting at OFF of Coda Parameter I). If nodes are invisible for complexity purposes, then there is no particular motivation for parameters such as the Coda Parameters (42). The fact that the default setting of Coda Parameter I is OFF (i.e. no LARYNGEAL node may be present in a coda) is then more or less accidental. If, however, nodes are visible for establishing complexity counts, and codas are (as we shall see in more detail in chapter 4) positions which are universally under pressure to reduce their complexity, then the default setting of Coda Parameter I falls out as a consequence of universal complexity considerations. In addition to this, including class nodes in the complexity count removes indeterminacy from our representations by allowing us to distinguish formally between (53a) and (53b). Under this approach, the representation in (53a) has a complexity of three, while that in (53b) has a complexity of only two. To avoid having two distinct representations of the same segment (such as the ones in (53)), one could then simply adopt a strategy based on Occam's Razor and assume that the representation with the lowest complexity which is compatible with the acoustic facts is to be chosen where more than one representation is available. The second strategy for determining segmental complexity would be to count class nodes as well as 'genuine' elements. This would have none of the disadvantages of the first approach and all of the advantages already indicated in the previous paragraph. Finally, we could count 'genuine' elements and those class nodes which are terminal nodes, that is, which have no daughters and which receive separate phonetic interpretation. Let us consider the implications o f this approach, using the RESONANCE node to illustrate the discussion. Consider the representations in (54). (54)

a.

χ

b.

I ROOT

χ I

ROOT

LAR

LAR

Η Η RES RES

·

V U M

58

[s]

One could, of course, argue that it is not the node which fails to be licensed but the elements (H and L) which it dominates. This would imply, however, that there is no justification for having a LARYNGEAL node in the first place. And yet, the evidence in favour of recognising such a node seems overwhelming (see e.g. Clements 1985, McCarthy 1988 and Kenstowicz 1994: chapters 4 and 9).

146

If we count the 'genuine' elements present in the representations in (54a) and (54b) as well as those class nodes which define a privative opposition, we arrive at the following result. (54a) has a complexity value of three (with h, H and U only being counted) and (54b) has a value of three as well (with h, H and RES being counted). This implies that coronals are as complex as their non-coronal congeners and should thus be found in the same environments as non-coronals - if complexity is a significant criterion at all. Convincing arguments in favour of interpreting complexity as a crucial and meaningful notion in phonology have been amassed in a good deal of GP work (especially KLV 1990, Harris 1990,1992,1994a and Harris & Lindsey 1995), so it seems wise to assume that complexity considerations have a role to play in this particular context as well. Assuming, then, that complexity matters, we need to check whether our complexity count of three for both (54a) and (54b) is compatible with the facts. Clearly, it is not. Coronals systematically differ from non-coronals in a number of ways (as summarised by Harris in (23)). For example, as stated in (23i), in many languages, a word-final consonant can only be a coronal. We know from Escure's environmental hierarchy (33) that word-final position is a weak environment, so we would expect coronals to be less complex than non-coronals. This, however, cannot be achieved if we count the RESONANCE node only where it has no daughters. If we were to count all elements and all nodes, though, we would get the desired result. (54a) would have a complexity of six, whereas (54b) would have only five. It seems, then, that the strategy which reflects the properties of certain segment types best, while simultaneously providing motivation for the Coda Parameters (42) and relating general simplicity considerations to the central notion of complexity is the second, under which all elements and class nodes are taken into account for determining the complexity of a segment. Returning from this general discussion of the issues involved in establishing the complexity of a segment to the potentially problematic governing relation between an onset occupied by [j] and a coda associated with [t] as shown in (52), we can now say the following. The complexity of [t] is four, which is greater than the complexity of [j], as [j] can muster only three. According to the Complexity Condition (19), the governor must be at least as complex as the governee, so the analysis as developed so far predicts that [tj] is not a possible coda-onset cluster. This comes as no surprise, since heterosyllabic clusters appear to be more preferred cross-linguistically if the coda is more sonorous than the following onset, to put it in non-GP terms. This point is expressed by Vennemann (1988: 40) in the Contact Law (55), which constitutes a slightly more perspicacious version of Murray & Vennemann's (1983) Syllable Contact Law. (55)

Contact Law A syllable contact A$B is the more preferred, the less the Consonantal Strength of the offset A and the greater the Consonantal Strength of the onset B; more precisely - the greater the characteristic difference CS(B)-CS(A) between the Consonantal Strength of Β and that of A.

If [tj] is not a possible coda-onset cluster, how is Adjektiv to be represented? As it happens, there is evidence pointing towards Adjektiv having internal analytic

147

morphology, which would mean that [t] and [j] are in separate domains and do not enter into a governing relation at all. The Latin preposition ad occurs as a separate preposition in several phrases which (like Adjektiv itself) are used in educated speech, e.g. ad acta, ad absurdum, ad calendas graecas, ad infinitum, ad libitum, etc. It can also be combined with German words, e.g. in ad eins ('re one'). In addition, there are Latinate terms used in syntax which also suggest that Ad- may be a separate morpheme. These include the pairs Präposition Adposition and Konjunkt - Adjunkt.59 So, one could argue Adjektiv should be treated in a similar way to the separable verbs ausweichen and losgehen and the noun Ablaß discussed earlier. The morphological structure of Adjektiv would then be as shown in (56). (56)

R

Ο I [ [χ I ?

Ν I χ I

a

Ο I χ I

t

Νλ

Ν Ο I I x ] [x j

I

X

X

I

I k

ε

Ο Ν I / \ X X X I V t i

Ο I X

Ν I X]]

Incidentally, we can now say whether the Complexity Condition would predict [pi] in Ablaß to be a possible coda-onset cluster. Counting all elements and nodes, for the reasons just spelt out, we find that [p] has a complexity of five (ROOT, RESONANCE, ?, h and U), while [1] has a complexity of only three (ROOT, RESONANCE and ?). It is clear, then, that [pi] is ruled out as a coda-onset cluster. Before moving on to a discussion of what the reduction in complexity caused by the loss of L means for the governing power of obstruents, particularly fricatives, I would like to deal with a potential problem for the analysis as presented so far. This problem arisesfromthe apparent existence of voiced obstruents in coda positions, that is, apparent counterexamples to my claim that coda positions are reserved exclusively for neutral (i.e. voiceless) obstruents. Consider the data set in (57), which consists of relatively recent borrowings, all typical Fremdwörter used almost exclusively in educated speech. With the possible exception of prägnant ('terse, pithy'), none of them require a gloss. Instead, I have given the corresponding form in the donor language(s), as shown in Drosdowski et al. 1976-1981. (57)

59

Lignin Magnesia Magnet Magnolie

[li'gniin] [ma'gneizia] [ma'gne:t] [ma'gno.'lis]

I owe these examples to Richard Wiese.

L: lignum ML: magnesia, Gr: magnesie L: magnés, Gr: mágnes F name: Magnol

148

Magnitude Dignitar prägnant Signal designieren Diagnose Prognose ignorant Stagnation Fragment Pragmatik Segment Pigment Magma Dogma

[magni'tu :da] [digni'ta:ç] [pre'gnant] [zi'gna:l] [dezi'gni:ran] [dia'gno:za] [pro'gno:za] [?i 'gnorant] [Jtagna'tsio:n] [fra'gment] [pra'gma:tik] [ze'gment] [pigment] ['magma] ['dogma]

L: mägnitudo F: dignitaire, ML: dignitärius F: prégnant, L: praegnâns F: signal, L: sïgnâlis L: designare F: diagnose, Gr: diagnosis L: prognosis, Gr: prognosis L: Ignöräns L: stagnare L: fragmentan Gr: pragmatikë L: segmentan L: pigmentum L: magma, Gr: mágma L: dogma, Gr: dògma

(Abbreviations: L = Latin, ML = Medieval Latin, Gr = Greek, F = French) The problem we are faced with is how to represent the syllable structure of these words. One possible approach to the noun Signal is illustrated in (58). (58)

R 0

N\

Ο

Ν

1 I\ I /\ x x x x x I ιI gI ηI \ a/ ζ

x

o n x

I I x I1

The trouble with the structure in (58) is that a voiced [g] occupies the coda position. According to my analysis, this should not be possible. Recall that only segments containing L are interpreted as voiced, and those, of course, are prevented from occupying this position by the settings for German of the Coda Parameters (42). The only conclusion to be drawn from this is that the [g] does not occupy a coda position but an onset. How, then, is the sequence [gn] to be syllabified? Treating it as a branching onset would seem the most obvious solution, but there are good reasons (to be discussed in 4.4.2) for rejecting this possibility. The only well-formed configuration available is one in which the [g] and the [n] are both attached to individual onset positions, separated from each other by an inaudible nuclear position. Recall that nuclear positions without segmental content (so-called 'empty nuclei') may remain inaudible if certain licensing requirements are met. I will show in 4.4.2 that this is the case in words such as Signal. So, until the necessary machinery has been introduced and the relevant

149

arguments have been put forward (all of which will be done in chapter 4), I shall simply suggest that the syllable structure of Signal and, mutatis mutandis, of the remaining words listed in (57) is that illustrated in (59).60 (59) Ο

Ν

Ο

l i l i χ χ I I I

χ

Ζ

I

g

R

«-

I

I O N I / \ I I X X X X X I \ / I

Ν χ

O

η

R N

a

1

I will return to these particular structures in chapter 4, where I will show that they also need to be invoked for native words such as Vennemann's variable items (e.g. eignen-, see 2.4.2.1.1). For now, let me take a closer look at the effects which the decomposition involved in FOD has on governing segments. Most of the subsection concerned with this will be devoted to fricatives, whose ability to govern is more likely to be compromised than that of plosives, due to their lower degree of complexity vis-à-vis plosives (which contain an additional element, ?).

3.4.3.2.3 Some devoiced obstruents can govern only homorganic nasals To see how the strength of a segment in terms of government is affected by the loss of L, it is necessary to consider configurations where a governing relation is known to hold between the FOD consonant and some other segment (strictly speaking, the skeletal positions they are attached to). It is in such configurations that segmental complexity is crucial. As we have already seen, configurations involving an alternating obstruent which governs another obstruent are exceedingly rare or, perhaps, non-existent (depending on how we deal with words such as Smaragd). This may well be an indication of the fact that the reduction in complexity which goes with FOD reduces the governing power of such a segment. Another obvious place to look for similar effects is in governing domains where the governee is a relatively complex sonorant, such as a nasal. However, it is worth reminding ourselves at this stage of the point made in 3.3.3 about nasals, namely that it is not entirely clear at present what exactly their composition should be. Still, as we shall see, the exact complexity values of nasals are not as important in the context of the discussion to follow as one may have expected. To see, then, how the application of FOD may reduce the governing power of the affected segment, consider, for example, the noun Bund ([bunt], 'league'). An underived 60

The details of the representation given here (including the licensing relations at the level of nuclear projection) will be further discussed in chapter 4.

150

form of this noun (to which FOD applies) is shown in (60a). This is contrasted with a suffixed form of the same root, where FOD does not apply, in (60b). (60)

RES

RES

In (60b) the plosive is clearly more complex than the nasal (having a complexity of six versus the three of the nasal).61 Even in (60a), there is still a complexity differential in favour of the plosive (four versus three of the nasal). So, no problems appear to arise where the governing onsets are occupied by plosives. Fricatives, however, may be problematic, since they are less complex than the corresponding stops. As argued by Harris (1990: 269), they lack the occlusion element ?, which is present in plosives. To see how fricatives perform under the same conditions as the oral stops in (60), consider the representations of Gans ([gans], 'goose') and Gänse (['genza], 'geese') shown in (61).

61

Recall lhat only nodes which are licensed by a particular position count towards the complexity of that position. So, only ROOT, Ν and ? can be included for the nasal. As will become clear in the course of the discussion, a RESONANCE node is probably not licensed in German coda nasals, so that it is not considered for complexity purposes.

151

b. ['genza]62 R Ο

Ν

O

Ν*

X X X g

RES

Ο

Ν

χ

χ

ε

RES

The governing onset in (61b) has a complexity of five, while the coda it governs has a complexity of only three. In (61a), by contrast, the complexity slope between the two corresponding positions is level, which, however, is no obstacle to the inter-constituent governing relation between the positions occupied by [n] and [s] respectively being established. Perhaps it is worth noting at this stage that Giegerich (1989: 6) assumes Gans to be homophonous with ganz ([gants], 'whole'). I do not actually share his intuitions on this point, but Giegerich is certainly right in observing that it is possible for a plosive to emerge in nasal + fricative and lateral + fricative clusters. This phenomenon of stop epenthesis or excrescence has been described and discussed by Ali et al. (1979), Dinnsen (1980), Fourakis & Port (1986) and Clements (1987), among others. Harris (1994a: 132f.) adopts the view that excrescence involves the spreading of the occlusion element ? from the nasal or lateral to the following fricative. As this element can be licensed in the governing onset, this kind of spreading is not just a matter of identification, but actually involves the acquisition of an element, so that the position targeted by spreading is strengthened. In the case of Gans, a realisation with excrescent [t], i.e. [gan's] would involve the governing onset gaining an extra element (?) and an additional node (a ROOT node to which the occlusion element is attached, so that an affricate-type segment is created). This results in a more favourable complexity slope than would be available in [gans], although, as already mentioned, it is still possible to establish a coda-onset governing relation involving [n] + [s], even where no excrescent [t] is present. The structure of the [n's] cluster is shown in (62), which is adapted form Harris (1994a: 133).

62

The details of the change in vowel quality (due to umlaut) need not concern us here. For a statement of the facts and an analysis in a different framework see e.g. Lodge 1989.

152

(62)

χ

We have seen that excrescent or epenthetic stops can improve the complexity slope, but that they are not essential for even devoiced fricatives to govern homorganic nasals. What, however, would happen if the nasal preceding the devoiced fricative did not share its place of articulation, that is, if it contained a RESONANCE node (and resonance element in the case of non-coronals) of its own? The relevant configuration, using an apparent [ms] cluster, as it occurs in the word Gems' ([gems], 'chamois'), as an example, is shown in (63).

RES ^

RES ·

U [m]

[s]

If [ms] in Gems ' really represents a coda-onset cluster, then there is a conflict with the Complexity Condition (19), as [m] has a complexity value of five and [s] of only three. The following prediction can be made on the basis of the discussion so far. If, due to a possible conflict with the Complexity Condition, an FOD fricative can govern a nasal only if that nasal identifies with its governor for its place of articulation, then FOD fricatives should occur exclusively in homorganic nasal + fricative clusters. Nasal + fricative clusters involving differing places of articulation should be reserved for one of the two following cases. Either there is an analytic boundary present, which separates the nasal from the fricative, so that no governing relation exists between the two, or the fricative has lexical H, which provides it with sufficient complexity (adding two) for the fricative to govern the non-homorganic nasal. To test this prediction, we need to examine non-homorganic nasal + fricative clusters, which are quite rare in German. A few, though, can be found, and these are listed in (64).

153

(64)

a. Wunsch Punsch Mensch

[vunj] [puni] [meni]

'wish' 'punch' "human being'

Ramsch

[rami]

'junk'

Hanf fünf Senf Genf

[hanf] [fvnf] [zenfj [genf]

Tiemp' 'five' 'mustard' 'Geneva'

manch Mönch Fench

[manç] [mœnç] [fenç]

'many' 'monk' '(type of) millet'

b. Gemse Bremse Tremse Wams Sims Bims

['gemza] ["bremza] ['tremza] [vams] [zims] [bims]

'chamois' "brake; horse-fly' 'cornflower' 'jacket' '(window) sill' 'pumice'

Bums Gesums Krimskrams

[bums] *bang' [ga'zums] 'fuss' [krimskrams] 'odds and ends'

If the nasals and the fricatives in (64) are actually adjacent, then we would expect no alternation to take place, as the fricative must have H to be able to govern. This is what we find in the majority of the words in (64a). With the exception of [f] in fünf (which can alternate with funfe, [ÎYnva]), none of these fricatives exhibit a voicing alternation. (Incidentally, the voicing alternation in fünf - ßnfe is not a problem, since the place element U present in the labio-dental fricative provides the necessary segmental complexity.) Looking at (64b), we do find an alternation between [z] and [s] in suffixed forms or, in the case of Gemse, in the pronunciation with apocopated schwa (similarly in Bremse and Tremse) or the variant form Gams ([gams]). Given the Complexity Condition (19) and my analysis of FOD, this can only mean that the alternating fricative does not actually govern the nasal. If thefricativedoes not govern the nasal, then the two positions cannot be adjacent, but have to be separated in some way.

154

Before investigating how this separation can be achieved, I would first like to take a look at non-homorganic nasal + stop clusters. Plosives are more complex than fricatives because they contain the additional element ?. So we would expect nasal + stop clusters to occur more freely than nasal + fricative clusters. Remarkably, though, the number of such clusters is extremely limited. A search of a reverse index dictionary of some 175, 000 entries (Muthmann 1991, which was also used for the data in (64)) and a machine-readable dictionary of similar proportions yielded only the following. (65)

a. Amt gesamt Samt Zimt Kumt Grumt

[?amt] [gs'zamt] [zamt] [tsimt] [kumt] [grumt]

'office'; MHG ambet 'whole'; MHG gesament 'velvet'; MHG samit 'cinnamon'; MHG zimmet 'horse-collar'; NHG Kummet, MHG komat 'aftermath'; NHG Grummet, MHG gruo(n)mät

b. fremd Hemd Öhmd verleumden

[fremt] [hemt] [?0:mt] [feç'bimdan]

'strange'; MHG fremede 'shirt'; MHG hemede 'aftermath'; MHG üemet, OHG uomät 'to slander'; NHG Leumund ('reputation')

As is clear from (65), all of these words can be related to earlier forms or even synchronic alternants containing a vowel between the nasal and the stop.63 One may speculate that the vowel present in the older forms has since been syncopated but has left traces of its presence behind. Although it is not clear at the moment what exactly these traces look like, the evidence seems to suggest that the nasal and the following stop are not adjacent even today.64 This means that the two are not in a governing relation. This is actually desirable, since, on closer inspection, it becomes apparent that a coronal plosive is only able to govern a bilabial nasal if it (the plosive) contains a laryngeal element. In that case, its complexity is six, which compares favourably with the nasal's five. A devoiced coronal plosive, which lacks a LARYNGEAL node altogether, however, can only muster a complexity value of four, and would thus be unable to govern a bilabial nasal. From what I have just suggested we could conclude that there are no non-homorganic nasal + stop clusters in German where the stop actually governs the nasal. In apparent non-homorganic nasal + stop clusters, it seems, the nasal does not occupy a coda position but an onset which is separated from the onset occupied by the plosive in some way. 63

All the etymological information in this section comes from Drosdowski et al. 1976-1981. See Harris 1994a (pp. 81f.) for a slightly more detailed discussion of similar cases in English, viz. the verbs hoax, coax and traipse, where the [s] behaves like an analytic suffix, as far as the phonology is concerned.

155

Genuine nasal + stop clusters, where the nasal is associated with a coda position governed by the plosive, are always homorganic in German. German is not at all unusual in permitting only such clusters. A very large number of unrelated languages exhibit precisely the same behaviour. These languages are sometimes referred to as 'Prince languages', because it was in Prince 1984 that their existence was first drawn to the attention of the phonological world (see also Goldsmith 1989 and Harris 1990 for further discussion). A true Prince language has Coda Parameter settings which permit maximally Ν and ? in the coda, so that the only attested coda-onset clusters are either geminates (where the coda itself is empty and derives its segmental content from sharing that of the governor) or nasal + consonant clusters (where the coda contains Ν and ?, but shares the RESONANCE node of the governor). German, as we have seen, is not a true Prince language, because it permits non-geminate obstruent clusters (e.g. in acht ([?axt], 'eight') or Luft ([luft], 'air')), but, as far as nasals are concerned, it does exhibit the hallmarks of a Prince language. If that is the case, then the putative nasal + fricative clusters in (64) must have a different explanation. They can no longer be interpreted as a configuration where a nasal with its own place element occupies a coda position which is governed by a fricative at a different place of articulation. Such a new explanation suggests itself when we consider earlier forms or contemporary foreign cognates of the words listed in (64). They either contain a vowel between the nasal and the fricative or there is evidence for an analytic structure involving a consonant-initial analytic suffix.65 This is confirmed in (66).66 (66)

65

66

a. Mensch

[meni]

OHG mennisco, mannisco

Ramsch

[rami]

French ramas, ramasser

Hanf Senf Genf

[hanf] [ζεηί] [genf]

MHG hanef, OHG haaf MHG senef French Genève

There are a few exceptions to this generalisation it seems, viz. Wunsch (MHG wünsch, OHG wunsc), finf (MHG vunv, vünv, OHG flnf, fünf) and Punsch (from English punch). It is worth noting, though, that these exceptions only involve configurations where the governing onset is associated with an expression which contains a resonance element (I or U). This resonance element provides sufficient complexity for the fricative to govern the coronal nasal anyway. This, however, does not solve the problem raised by the fact that the coronal nasal must contain a RESONANCE node of its own, which is in conflict with my claim that coda nasals in German can license no elements or nodes other than ROOT, Ν and ?. At present, I can offer no explanation for this. See Fleischer & Barz 1992 (pp. 167f.) for a discussion of the suffix -s which supports my claims about the morphological structure of the second group of nouns in (66b).

156

manch Mönch

[manç] [mœnç]

MHG manee, manig MHG mönnich

Fench

[fenç]

NHG Fennich, MHG phenich, OHG phenih

['gemza] [vams] [zims] [bims] Cbremza] ['tremza] [bums] [ga'zums] [krimskrams]

MHG geme je Old French wambais MHG simej MHG bümej, OHG bumij MLG premese·, OHG brimissa MLG trem(e)se [[Bum]s] [[[Ge][summ(e)]]s] reduplicated form of [[Kram]s]

b. Gemse Wams Sims Bims Bremse Tremse Bums Gesums Krimskrams

So, it appears to be the case that non-homorganic nasal + fricative clusters are rare, but this is not because the fricatives cannot govern but because of a special property of the German coda (a property which, to some extent, it shares with codas in true Prince languages). The fact that non-homorganic nasal + stop clusters are at least as rare supports this interpretation. So, we can learn only very little about what the reduction in complexity associated with FOD means for the governing power of fricatives by investigating apparent nasal + fricative clusters. The rarity, or possibly complete absence, of alternating obstruents (both stops and fricatives) after another obstruent is probably more revealing. It strongly suggests that such configurations would be ill-formed due to the low degree of complexity, and thus reduced governing power, of a devoiced segment.

3.5 Conclusion In this chapter, I have presented the first part of my GP analysis of FOD in German. After an introduction to some of the central notions of GP and its representational apparatus, I have proposed segmental representations for both obstruents and sonorants which are relevant to the discussion of FOD. Focusing on what the changes are which take place when an obstruent undergoes FOD, I have first of all considered the issue of how voicing is manifested both in acoustic and articulatory terms. The conclusion drawn from my survey of possible cues has been that no reliable and universally applicable phonetic cues to the voicing contrast can be pinpointed and that the argumentation has to be based on phonological considerations alone. I have examined several hypotheses as to what FOD might involve in the GP framework at a fairly superficial level and found that, at first blush, treating FOD as a

157

reduction phenomenon involving the loss of lexical L is more promising than the other options available. I have put this hypothesis to the test by examining its implications and verifying the predictions which can be made on the basis of it. The most important among these are that FOD is a phonological weakening process and that the application of FOD does not lead to complete neutralisation. Interpreting FOD as a weakening process brings it into line with other, more obvious, weakening processes such as spirantisation and deletion, which are particularly common in word-final position, that is, in the most well-known FOD environment. My claim that the application of FOD does not lead to complete neutralisation receives support both from experimental studies and from phonological events such as the spirantisation of velars, both of which will be discussed in more detail in chapter 5. I have also demonstrated that what other researchers (e.g. Rubach (1990)) have analysed as syllable-final devoicing does not involve the application of any dynamic phonological process at all. Instead, it falls out from parameter settings which constrain the range of segmental material which can be licensed in coda positions, so that codas can accommodate only segments which lack a LARYNGEAL node and are thus interpreted as voiceless. On the whole, the predictions made by my account of FOD as phonological weakening are borne out by the facts. My analysis of FOD as a reduction process, therefore, seems reasonably well supported. Having dealt with the issue of what FOD is, I will address the problem of where it occurs in the next chapter.

4. Where Does Final Obstruent Devoicing Occur? A Government Phonology Approach

4.1 Introduction This chapter contains the second part of my Government Phonology analysis of FOD. It tackles the issue of where FOD occurs. As shown in chapter 2, this problem has received rather more attention in the literature over the past two decades and a half than the question of what FOD is. A variety of solutions have been proposed, all of which are beset by more or less serious problems. I will argue that a more successful treatment of FOD has to make reference to empty categories, specifically, to empty nuclear positions, as well as the notion of licensing. In what follows I will concentrate on true voicing alternations (i.e. FOD proper) only, since the question of how to handle distributional restrictions involving voicing has already been dealt with in chapter 3.1 begin by examining the most straightforward FOD environments, from which I move on to a discussion of Vennemann's variable items.

4.2 Straightforward FOD environments 4.2.1 Word-final and term-final in compounds Before going into the details of where FOD occurs, let me provide a brief reminder of the sort of facts to be accounted for, by repeating part of the survey of FOD forms from chapter 1 here, for convenience. In chapter 1,1 observed that FOD is triggered wordfinally, before certain suffixes and at the end of a term within a compound. I shall take the simplest cases first, i.e. those where FOD applies word-finally. The relevant data are contained in list (1). Recall that, in this list, + represents any kind of morphological boundary, not just a non-analytic one. (!) FOD in word-final position a. Native words Dieb Stab Lob Laub Lied

[di:p] [Jtaip] [lo:p] [laup] [li:t]

['di:b+a] ['Jtaib+a] [lo:b+9] [laub+a] [liid+u]

•thief •staff 'praise' 'foliage' 'song'

Rad Tod Kleid Sieg Tag Trog Zweig brav fies Gas Los reis' Haus

[nut] [tort] [klait] [zi:k] [ta:k] [tro:k] [tsvaik] [braifj [fi:s] [ga:s] [lois] [rais] [haus]

['ra:d+9] ['to:d+s] ['klaid+B] ['ziig+a] ['taig+a] ['troig+a] ['tsvaig+s] ['bra:v+9] [Ti:z+3] ['ga:z+a] ['raiz+a] ['hauz+a]

•wheel* 'death' 'dress' 'victory' 'day' 'trough' 'twig' 'good; well-behaved' 'nasty' 'gas' 'lottery ticket' '(I) travel' 'house'

ebb' Kladd' jobb' egg' König

[?ερ] [klat] [fek] [•k0:niç]

[?eb+3] fklad+a] ['dpb+a] [Teg+9] ['k0:nig+a]

'(I) ebb' 'notebook' '(I) do a temporary job' '(I) harrow' •king'

halb Wald Balg Calw Hals Korb Bord Berg Kurv' Vers Hemd Gems' Bund Gans

[halp] [valt] [balk] [kalf] [hals] [koçp] [boçt] [beçk] [kuçf] [feçs] [hemt] [gems] [bunt] [gans]

[halb+9] [Vald+s] ['balg+9] ['kalv+B] [halz+s] [koçb+a] ['boçd+a] [•beçg+a] f'kuçv+a] [feçz+s] ['hemd+a] ['gemz+s] ['bund+a] ï'genz+s]

•half 'forest' •brat' (place name) 'neck' •basket' 'shelf 'mountain' "bend; curve' 'stanza' 'shirt' 'chamois' 'league' 'goose'

[sma'rakd+a] ['me:kd+3]

'emerald' 'maid'

[D3=»p]

Smaragd [smatakt] Magd [ma:kt]

[ΊΟ:Ζ+9]

160

b. Fremdwörter Cherub Snob rapid Monolog Verb Ford Chirurg kursiv konfus

fçeirup] [snap] [ra'piit] [monoloik] [veçp] [fo?t] [çi'ruçk] [kuç'ziif] [kon'fuis]

[çe:ru'bi:n9n] [sno'bismus] [ra'piida] [mono'lojgs] ['veçbgn] —

[çi'ruçgan] [kuç'zi:va] [kon'fuiza]

'cherub' 'snob' 'rapid' 'monologue' 'verb' 'Ford car' 'surgeon' italic' 'confused'

Focusing on the most common configurations, that is, the majority of the native words listed in (la), we find that there are two predominant structures involved here. They are shown in (2). (2)

a. Branching rhyme

b. Branching nucleus

R

R ι 1 Ν Χ*- χ I I {?} {ρ} {1} m {η} {k} {f} {S}

/ \X

X

\ a/

0 Ν 1 1 χ X 1 {ρ} {t} {k} {f} {s}

As already observed in earlier chapters, in the vast majority of cases, an FOD obstruent is preceded by either a long vowel or diphthong, or by orthographic I, r or n. Turning to what follows the FOD obstruent, we find greater consistency. In all cases in (1), the FOD obstruent is apparently word-final, that is, followed by a word boundary. GP, however, makes no provision for consonants in absolute word-final position, which is why, in this framework, the existence of such segments is only apparent. A well-formed phonological representation consists of O-R sequences, so the word-final position is actually a nucleus (as illustrated in (2)), to be precise, a parametrically licensed domain-final empty nuclear position. Although these positions have already been referred to in 3.2.1, they merit further discussion here. Parametrically licensed domain-final empty nuclear positions, as their name suggests, are empty nuclear positions which occur at the end of an analytic domain in languages where the relevant parameter is set at ON. These positions, or rather, the parameter which

161

controls them, have an important role to play in accounting for the special properties of apparently word-final consonants as well as in the typological classification of the world's languages. This potential, however, can only be fully realised owing to the fact that GP contains a principle known as the Coda Licensing Principle (see especially Kaye 1990b). To understand the function of this principle in the present context, consider the German word Bus ([bus], TJUS'). This could be represented as shown in (3).

X X X b

u

s

The Coda Licensing Principle, quoted from Kaye 1990b (p. 311) in (4), however, rules this representation out. (4)

Coda Licensing Principle Post-nuclear rhymal positions must be licensed by a following onset.

The coda position occupied by the [s] in (3) is not licensed, as there is no following onset. Since the [s] cannot occupy a coda position, it must necessarily be associated with an onset position instead. Onset positions, in tum, can be interpreted only if they are licensed by a following nucleus, which is stated in the Onset Licensing Principle (5), first proposed by Harris (1992: 380). (5)

Onset Licensing Principle An onset head position must be licensed by a nuclear position.

Therefore, the only well-formed alternative to the ill-formed syllable structure in (3) can be that shown in (6). (6)

Ο l χ I b

Ν i

Ο l

χ I u

Ν i

χ I

χ

s

The Coda Licensing Principle, working in tandem with the Onset Licensing Principle, then, stipulates that there are no non-nuclear positions immediately preceding a domain boundary, that is, all consonants which appear to be at the end of a word are actually followed by a domain-final empty nuclear position. This is true universally. Consequently, the occurrence of word-final consonants depends on whether a language

162

parametrically licenses domain-final empty nuclear positions. The languages of the world can be divided into two groups depending on whether they license such positions or not. (Further divisions are of course possible and necessary, but this is the one relevant to our discussion). The most clear-cut cases are at both ends of the spectrum. Languages exhibiting words which end in consonant clusters, such as help in English or Bund in German, license such positions. By contrast, languages which do not permit word-final consonants at all, such as Italian or Desano (Eastern Tucanoan; Colombia, Brazil), for example, do not. Languages which have only single word-final consonants (such as Korean or European Portuguese) require a more detailed analysis to decide whether the empty nuclear position is licensed by parameter setting or by proper government (see (44) and (45)). The Final Empty Nucleus Parameter in (7) captures this. Its default setting is OFF (which is also the setting for Italian and Desano), but it is set at ON for German (as well as for languages such as English and French). (7)

Final Empty Nucleus Parameter A domain-final empty nuclear position is licensed: ON/[OFF]

Coda Licensing and Onset Licensing, as already hinted at, can, however, not only contribute to more appropriate language typology, they also shed light on the special properties of apparently word-final consonants. A number of researchers have observed that, with regard to stress assignment, such consonants do not seem to 'count' (see e.g. Hayes 1982). Furthermore, they fail to trigger so-called closed-syllable shortening (e.g. Myers 1987) and they have a tendency to contravene otherwise general sonority sequencing generalisations (e.g. Levin 1985). Apart from this, as pointed out by Lombardi (1991: vii), in many languages, word-final consonants do not undergo laryngeal neutralisation, although word-internal syllable-final consonants do. If wordfinal consonants are interpreted as onsets rather than codas (which, as we have seen, is what Coda Licensing and Onset Licensing require), then these facts find a straightforward explanation (see Harris 1992 for further details). To sum up so far, I have observed that FOD applies whenever an FOD segment immediately precedes a domain-final empty nuclear position. However, I have restricted the discussion up to this point to those words where the affected obstruent is preceded either by a long vowel or diphthong, or by a sonorant consonant (or its vocalised reflex in the case of /r/). It appears that this, unlike the presence of a domain-final empty nuclear position, is not a sine qua non for the application of FOD, though. FOD can apply equally well to forms with a preceding short vowel (e.g. König, Cherub, Snob etc.) or a preceding obstruent. Two things are, however, worth noting. The fact that in the vast majority of cases it is a long vowel which precedes the FOD obstruent is due to what one can probably only describe as a historical accident (see

163

1.2.2.2 for a more detailed discussion), that is, a diachronic lengthening process which affected stressed vowels preceding voiced consonants.1 The second point is that, in the rare cases where the FOD obstruent is preceded by another obstruent or a non-homorganic nasal, e.g. in Magd, Jagd or Smaragd and Hemd or fremd (see 1.2.2.2 for details), there is evidence which suggests that a licensed empty nuclear position or, perhaps, an analytic boundary precedes the FOD obstruent.2 It is, for example, likely that Jagd began life as an analytic form, with a suffix containing a coronal stop. The fact that this stop is not part of the verb stem jag- (cf. jag-en, ['jaigan], 'to hunt') supports this view, which is further strengthened by the fact that the (now unproductive) deverbal feminine suffixes -de, -te and -d are all identifiable as having played a role in word formation in the past. The following sets of nouns (from Fleischer & Barz 1992: 198f.) bear witness to this. The fourth column in (8) shows the verb stem from which the noun in the first column is derived. Jagd Mahd

[ja:kt] [ma:t]

•hunt' 'mowing'

Blüte Freite

[blyita] ['fra ita]

'flowering' blüh'search for a wife' frei-

Freude Beschwerde Zierde Gelübde

['firaida] [ba'/ve:çda] ['tsi:çda] [ga'lypda]

'joy' 'complaint' 'decoration' 'vow'

jagmäh-

sich freusich beschwerziergelob-

'to hunt' 'to mow' 'to flower' 'to court' 'to be glad' 'to complain' 'to adorn' 'to vow'

It may be true that the noun is no longer analytic in terms of synchronic morphology because the suffix involved is no longer productive, but the phonological consequences of the original analytic structure may still be in place. So, from a phonological point of view, we may still be dealing with a configuration such as that set out in (9),3 where [k] and [t], as elsewhere in this book, reflect the surface pronunciation, not the underlying segments (which, of course, are voiced).

1

2

3

Interestingly, this lengthening process did not take place in Dutch (another language with FOD), so that voicing alternations after short stressed vowels occur much more commonly than they do in German. This fact further supports my view that the length of the preceding vowel is essentially independent of FOD itself. What matters is what comes after the FOD obstruent. I am grateful to Jan Kooij for drawing my attention to this property of Dutch. As we saw in chapter 3, FOD is best interpreted as segmental reduction. Segmental reduction, however, affects a segments complexity and thus its ability to govern (see 3.4.3.2.3), so that we would not normally expect a neutral obstruent (with reduced governing power) to be able to govern another obstruent. The MHG forms jaget or jagat suggest that the original form may have been [[jag]et] or [[jag]at], with a vowel-initial suffix. This vowel has since been lost, and the NHG form may be that shown in (9).

164

(9)

Ο I [[ χ I j

Ν

Ο

Ν

/ Χ χ χ

I χ

I χ]

N

a

/

I

k

Ν χ

I χ]

I t

Although it may not be possible to resolve all the questions posed by structurally unusual and rare words such as Magd, Jagd or Smaragd by means of the present analysis, the majority can be accounted for. Incidentally, I am not alone in assuming a morphological structure such as that shown in (9) for Jagd. Giegerich (1989: 55) argues that the -d is a level 2 suffix, a view which is shared by Yu (1992b: 172). Wiese (1988: 146) seems to agree with this interpretation as well, although he does not say on which level -d is to be attached. Kloeke (1982a: 31) also suggests that there is a morpheme boundary present in Jagd. Be that as it may, these somewhat problematic items, together with the more straightforward ones in (1), suggest the hypothesis in (10). (10)

FOD applies preceding a licensed domain-final empty nuclear position.

The remainder of this chapter will be devoted to refining and justifying (10). It certainly appears to hold also for compounds, where each term consists of an analytic domain of its own. In other words, a two-term compound (such as those shown in (11)) has the structure [[A][B]], with FOD applying to an obstruent before each domain-final empty nucleus. (11)

a. Nouns Leib-wächter Leib-eigener Lob-gesang Farb-stoff Schreib-tisch Wald-brand Wald-ameise Rad-nabe Geld-beutel Bild-hauer Weg-weiser Aug-apfel Berg-steiger Berg-amt

flaipveçte] [laipPaiganß] [lo:pg9zaq] [Îaçp/tof] ['JraiptiJ] ['valtbrant] ['valt?a:maiz9] ['raitnaibs] ['geltboitsl] ['bilthaue] [Ve:kvaizB] ['?auk?apfsl] [•beçk/taigr?] ['beçk?amt]

'body guard' 'serf 'song of praise' 'colouring' 'desk' 'forest fire' 'red ant' 'hub' 'purse' 'sculptor' 'signpost' 'eye-ball' 'mountaineer' 'Mining Office'

165

Zweig-stelle Haus-tier Haus-arzt Hals-band

['tsvaik/teb] [hausti:ç] ['haus?açtst] f'halsbant]

"branch' 'pet' 'family doctor' 'dog collar'

b. Adjectives lob-abhängig ['lo:p?apheqiç] farb-echt [façp?eçt] schlag-empfindlich [ 'Jla :k?empfι ntl ι ç] flug-fertig ['flu:kfeçtiç] haus-ähnlich ['haus?e:nliç] geld-orientiert ['gelt?orÎ3nti:çt] trag-föhig ['tra:kfe:iç] leid-voll [laitfol] schlag-artig ['Jla:k?açtiç] sieb-förmig ['zi:pfœçmiç] staub-frei ['Jtaupfrai] wald-arm ['valt?açm] halb-leer [halpleiç]

'dependent on praise' 'colour-fast' 'shock-sensitive' 'ready for flying' Tiouse-like' 'money-oriented' 'strong' 'sorrowful' 'sudden' 'sieve-shaped' 'dust-free' 'sparsely wooded' "half-empty'

c. Verbs blind-schreiben lob-preisen gesund-pflegen übrig-bleiben wund-liegen wund-arbeiten stand-halten

['blint/raiban] [loippraizan] [ga'zuntpfleigsn] ['yibriçblaiban] ['vuntliigan] ['vuntîaçbaitan] ['Jtanthaltan]

'to touch-type' 'to praise highly' 'to nurse (someone) back to health' 'to be left over' 'to develop a bedsore' 'to work (one's fingers) to the bone' 'to stand firm'

4.2.2 Before certain suffixes So far, (10) makes the right predictions. What about suffixes which trigger FOD in a stem-final obstruent? One could argue that these suffixes involve two domains, that is, a structure such as [[A]B], with A representing the root and Β the suffix. Given that the relevant suffixes do not affect stress-assignment within domain A and cannot bear primary stress (that is, they are stress-neutral), that they are usually productive and do not tend to exhibit lexical selectivity, this assumption is reasonable. In other words, I propose

166

to treat the suffixes in (12) as analytic. Then it is possible to extend the analysis as presented so far to these suffixed forms as well. FOD applies before a licensed domainfinal empty nuclear position in all cases. (12)

a. Nominal (i) Derivational Suffix

Example word

-1er -heit -keit -ling -nis -sal -sel -schaft -tum -chen -lein -de

Häusler Kindheit Farbigkeit Liebling Ergebnis Labsal Geschreibsel Liebschaft Herzogtum Hündchen Äuglein Gelübde

[hoislB] [•kinthait] [façbiçkait] [liipliq] [?eç'ge:pnis] [laipzail] [gs'/raipssl] [liipjaft] ['heçtso:ktu:m] ['hYntçan] [•?3iklain] [ga'lypda]

'cottager' 'childhood' 'colourflilness' 'darling' 'result' 'refreshment' 'scribbling' 'love affair' 'dukedom' 'little dog' 'little eye' 'vow'

-bold -werk -gut

Tugendbold Laubwerk Treibgut

['tu:gantbolt] [laupveçk] î'traipgu:t]

'paragon of virtue' 'foliage' 'flotsam'

['?u:çlaups]

'holiday', gen. sg.

(ii) Inflectional Suffix

Example word

-s

Urlaubs

b. Adjectival/Adverbial (derivational suffixes only) Suffix

Example word

-bar -los -haft -lieh

lösbar farblos glaubhaft kindlich

[l0:sbaç] ['façplo:s] ['glauphaft] ['kintliç]

'solvable' 'colourless' "believable' 'child-like'

167

-sam -kundig -mäßig -wert -lustig -fest -wärts -lings

kleidsam schreibkundig bildmäßig preiswert schreiblustig schlagfest abwärts blindlings

['klaitza:m] ['Jraipkundiç] [•biltme:siç] ['praisveiçt] ['/raiplustiç] [•Jlaïkfest] [7apveçts] ['blintliqs]

•becoming' 'literate' "by means of a picture' 'good value for money' "keen to write' 'shock resistant' 'down' Tjlindly'

c. Verbal (inflectional suffixes only) Suffix

Example word

-t -st -te

schreibt schreibst liebte

['Jraipt] ['jraipst] [liipta]

'(he/she/it) writes' '(you, sg.) write' '(he/she/it) loved'

Now, what exactly is it that distinguishes the analytic suffixes in (12), which trigger FOD, from the suffixes in (13),4 which don't? (13)

a. Nominal (i) Derivational

4

Suffix

Example word

-erei -el -er -icht -ung -e -in

Dieberei Hebel Schreiber Weidicht Färbung Binde Hündin

[diiba'rai] f'heibsl] ['/taita] ['vaidiçt] ['feçbuq] ['binda] [TiYndin]

'thieving' 'lever' 'scribe' 'place where willows grow' 'colouring' •band' •bitch'

As already pointed out in 1.2.2.2, not all German suffixes which fail to trigger FOD are listed here. Only those are included which can be attached to roots exhibiting voicing alternations. That is, the roots concerned take both FOD-triggering suffixes (or form part of a compound) and suffixes which do not trigger FOD. The suffix -ant, for example, as in amüsant ('amusing'), which is in fact a nontriggering suffix, was omitted because none of the roots which can take -ant can also be combined with suffixes which would trigger FOD, that is, they can never exhibit any voicing alternation at all.

168

-ur -ation -ik -ität

Glasur Delegation Motivik Naivität

[gla'zu:ç] [delega'tsio:n] [mo'tiivik] [na?i:vi'te:t]

'glaze' 'delegation' 'motif 'naivety'

['gra:bas] [liiban] ['heçts0:g3] [ΊΙΟΙΖΒ] ['ho I zen]

'grave', gen. sg. 'loved ones' 'dukes' 'houses' 'houses', dat. pl.

(ii) Inflectional

b.

Suffix

Example word

-es -en -e -er -ern

Grabes Lieben Herzöge Häuser Häusern

Adjectival (derivational suffixes listed below, inflectional suffixes mostly identical with those of nouns)

Suffix

Example word

-ig -isch -en -ern

leidig kindisch seiden gläsern

[laidiç] [kindii] [ za ι dsn] î'gleizen]

'tiresome' 'childish' 'silken' '(made of) glass'

-iv -ÖS

impulsiv nervös

[?impul'zi:f] [neç'v0:s]

'impulsive' 'nervous'

c. Verbal (inflectional suffixes only) Suffix

Example word

-en -e

schreiben schreibe

['/ra iban] ['/ra iba]

'to write' '(I) write'

-ieren

rasieren

[ra'zi:ran]

'to shave'

It is obvious that the suffixes in (13) are all vowel-initial, whereas all of those in (12) are consonant-initial. Before dealing with this point, let me first see how the hypothesis that

169

FOD applies before a licensed domain-final empty nuclear position stands up to the words in (13). Consider the representation of the verb schreiben ('to write') shown in (14). The infinitive suffix -en is stress-neutral, exhibits no lexical selectivity and is even used productively, so there can be little doubt that it is analytic - hence the bracketing in (14). (14)

O / \ [[χ χ I I / r

N

O

N

/ \ X X I I a I

I X I ρ

I X]

Ν I X I a

X I η

Ν I X]

The FOD obstruent immediately precedes a domain-final empty nuclear position in (14). So, on the first cycle, when only the domain containing the stem is available, FOD will apply. On the second cycle, however, the final empty nucleus which triggered FOD on thefirstcycle is deleted under reduction/the OCP (see 3.2.2). Recall that the suffix-initial empty onset (not actually shown in (14) due to its predictability) has no effect on the application ofreduction/theOCP, as it dominates no skeletal position. The configuration on the second cycle is then as shown in (15). Ο

(15)

[

/ \ x x

J

Ν x

/ \ x a

x i

Ο

Ν

Ο

ι

x

ι x

ι x

b/p

a

η

Ν ]

ι

The nuclear position following the FOD obstruent is now filled, so, at this stage, FOD should be blocked, which is precisely what the actual pronunciation ['/raiban] suggests. However, FOD has already applied, on the first cycle. The laryngeal element L and the LARYNGEAL node have already been delinked. Reattaching them again on the next cycle would amount to a Duke of York gambit (see Pullum 1976). As already argued in 2.4.4.3, an analysis which can do without this strategy must be considered preferable to one which needs to employ Duke of York derivations. So, if the Duke of York gambit which involves reattaching the delinked L and LARYNGEAL node is to be rejected, then two options are open to us. Either the analysis of FOD presented in chapter 3 has to be changed in some way or the suffix -en must be interpreted as non-analytic. A nonanalytic suffix -en would yield the following structure. Recall that the formative boundary + is invisible to the phonology in GP, that is, it is treated as if it did not exist.

170 (16)

O

Ν

/ \ [χ χ I

I

;

r

/ \X

χ I 1 a

I 1 ι

0 Ν 1 1 χ + X I I 1 1 b 3

0

Ν

I

1

X ι 1 η

X]

FOD would be blocked because the FOD obstruent is followed by a filled, as opposed to a licensed empty, nuclear position, so that the correct output is generated. The trouble is that not just -en would have to be treated as non-analytic, but all other vowel-initial suffixes in German (including those listed in (13)) as well. This is problematic since there are strong arguments against many of the vowel-initial suffixes being interpreted as non-analytic. Recall that analytic suffixes are typically stress-neutral, exhibit no lexical selectivity and are used productively. The infinitive suffix -en, for example, has all these properties, so it appears to be analytic. Authors working in a Lexical Phonologyframework(e.g. Wiese 1988, Hall 1989a and Giegerich 1989) acknowledge this by assigning regular inflection (to which -en belongs) to level 3. Other suffixes which are attached at this level (and which need to be treated as analytic in GP) include the present indicative verb suffixes -e (1st sg.) and -t (3rd sg.) as well as the regular noun plural suffix -n (see Hall, op. cit.).5 Suffixes which are stress-neutral and may be used productively, but exhibit some lexical selectivity are usually assigned to level 2 in Lexical Phonology work. Hall (op. cit.), for example, lists the (mainly deverbal) derivational suffixes -nis, -ung, -ling, -ig and -lich among the level 2 affixes. According to Hall (who follows Giegerich 1985, 1987 and Wiese 1988), class II derivation6 and compounding take place at level 2. Considering that the level 2 affixes of Lexical Phonology are stress-neutral and productive (see Fleischer & Barz 1992), they must be treated as analytic in GP, in spite of the fact that, as Giegerich (1989: 8) observes, at least some of them exhibit lexical selectivity in that they make reference to the stress pattern of the root to determine which roots they can attach to. Level 1 affixes, by contrast, are not stress-neutral, which is why Lexical Phonology accounts assume them to fall into the domain of the rules assigning stress. The general consensus appears to be that stress assignment takes place on level 1, so class I derivation and irregular inflection affecting stress belong to this level. It seems most appropriate to capture these facts in GP by interpreting the relevant structures as non-analytic. That is, class I derivation and irregular inflection are invisible to the phonology, so that words

5

The examples Hall (1989a: 809) cites here are bring-e (['briga], '(I) bring'), bring-t ([bnijt], '(he/she/it) brings') and Hügel-η (['hyigaln], "hills'). The level 2 examples he provides are Erleb-nis ([?eç'le:pnis], 'experience'), Les-ung ([leizurj], 'reading'), Lehr-ling (['leiçliq], 'apprentice'), wonn-ig (['voniç], 'delightful') and freund-lich ([frointliç], 'friendly'). Wiese (1988: 150) distinguishes between class I affixes and class II affixes by observing that the former occur only between the latter and the root, but never vice versa. Furthermore, class I affixes can bear the main stress of a word, while class II affixes do not affect stress assignment and cannot carry main stress.

171

such asAktiv-isí ([?akti'vist], 'activist') and Schreib-erei ([Jraiba'rai], 'writing') have the same status as morphologically simple words such as Mann ([man], 'man') or neu ([noi], 'new').7 The upshot is, then, that there are indeed some vowel-initial suffixes in German which require the kind of structure illustrated in (16), but infinitival -en is not one of them. Instead -en (and a number of other vowel-initial suffixes) need to be treated as analytic. These arguments appear to force the conclusion that it is the analysis of FOD proposed in chapter 3 which is incorrect. In that chapter, I argued that L is involved in some kind of reduction event, and this claim appears to be well enough supported. What is completely unsupported, however, is the assumption that FOD is necessarily a dynamic process, in the sense that it involves a change in the segmental representation. An alternative to treating FOD as a process is considering it as a matter of phonological interpretation, specifically of autosegmental licensing.

4.3 FOD as an instance of autosegmental licensing To show how the notion of autosegmental licensing (due to Goldsmith (1989, 1990: chapter 3)), can be invoked for the purposes of an analysis of FOD, we need to consider first of all what is meant by autosegmental licensing and how it can be accommodated in the GP framework. Goldsmith's original idea was that no (privative) feature could be interpreted as part of a representation, unless its presence was sanctioned by a licenser. This is captured in the Licensing Criterion (17), quoted from Goldsmith 1989 (p. 150). (17)

Licensing Criterion Each distinctive feature in a representation must be licensed by its closest licenser [...]. Each licenser may license no more than one occurrence of each feature.

Goldsmith assumes the hierarchical syllable structure illustrated in (18), where the syllable node and the coda node function as licensers. The notion of'closest licenser' is then straightforward: it is whichever node is closer to the position in question.

7

As we shall see in 4.4.2.6, these categorisations are only rough guidelines and will have to be further refined.

172

(18)

σ

o

I

n

o = onset r = rhyme η = nucleus c = coda e

I

I

X X X Goldsmith argues that the set of features which can be licensed by the coda node is a subset of the features which can be licensed by the syllable node. In true Prince languages, for example, the coda node cannot license a separate place feature or any other feature, except for [nasal], which accounts for the fact that only geminates or nasal + homorganic stop clusters may occur in such languages. The idea that separate licensing is required for melodic material found its way from Goldsmith's work into GP via Harris 1990 (especially p. 284) and Shohei Yoshida 1990, which, in turn formed the basis for Brockhaus 1992b, where the term 'autosegmental licensing' was first used in the GP framework. The notion has since been greatly developed by Harris (see especially Harris 1992). Building on work in other phonological theories (e.g. Selkirk 1978, Nespor & Vogel 1986, Goldsmith 1989, 1990), Harris draws a distinction between prosodie licensing and autosegmental licensing. The former (p-licensing for short) sanctions the presence of positions at different levels of projection from the skeletal tier upwards, while the latter (a-licensing) determines the melodic content of a particular position. In the context of a-licensing, the notion of a-licensing potential, as defined in (19), plays an important role (Harris 1992: 384). (19)

A-licensing potential The a-licensing potential of a skeletal position refers to its ability either (a) to directly a-license a melodic expression, or (b) to confer a-licensing potential on another position.

The a-licensing potential of a particular position depends on its place in the prosodie hierarchy, as stated in the Licensing Inheritance Principle (20). (20)

Licensing Inheritance (Harris 1992: 384) A licensed position inherits its a-licensing potential from its licenser.

The specific claim made in (20) is that how much a-licensing potential a position has depends on how far removed from the head of the domain it is along the licensing path. We saw in 3.2.1 that licensing is the motor which drives phonology and that all positions within a domain, bar the head of that domain, have to be licensed. Following the licensing relations from the head of the domain to any position within the domain amounts to tracing a licensing path. To see how this works, consider the representation in (21), which is slightly adapted from Harris 1992 (p. 384). It is the representation of the English word tawdry, as pronounced in RP, and it illustrates how the head of the

173

domain p-licenses other positions within that domain. (21) differs from the usual GP-representations, such as (16), for example, in that constituency is shown by means of brackets rather than arboreally. This is done in an effort to represent the licensing relations as clearly as possible.

(21)

[x,]0

I

t

[x2

X3]N

\ / o

[X 4

I

d

XJO

I

J

[XJN

I ι

What we see in (21) are licensing relations between the following pairs of positions (licenser first): x2 and x b x2 and x3, x2 and x^, x¿ and x4, and x4 and x5. Two of the five represent cases of intra-constituent licensing relations between adjacent positions. They are x2 and x3 (branching nucleus), and X4 and x5 (branching onset). The licensing relation between x2 and x,, by contrast, is an inter-constituent licensing relation, where the nucleus licenses the immediately preceding onset. The licensing relation between x2 and Xe is established at a level of nuclear projection. Positions x¿ and x4 must also be assumed to be in a licensing relation, as a nucleus licenses the preceding onset head. The problem with this particular licensing relation is that the two positions involved are not actually adjacent. This is a point of which GP researchers are well aware, but, as far as I know, no completely satisfactory solution has as yet been proposed. Heads of phonological domains are unlicensed within that domain but, with the exception of the head of the morphological domain itself, receive their own licensing from outside the domain of which they are the head. In (21), for example, x4 is licensed by x6. The latter position, x6, is unlicensed within the domain defined by the licensing relation between x4 and x^ but receives its licensing from x2. The only position which is unlicensed throughout the morphological domain of which this word consists is x2. It is thus the head of this domain and, as such, bears the main stress. If we trace the licensing path from x2, the head of the morphological domain, to, say, x5, we can see (simply by counting the arrows, each of which represents one 'dose' of p-licensing) that x5 is at three removes from the head, whereas, X! is at no more than one remove. If we follow Harris in assuming that each remove from the head results in depletion of the a-licensing potential of a position, then we come to the conclusion that the a-licensing potential of x5 is smaller than that of x,.8 In other words, we expect x5 to

8

Note that Harris (1992: 386) actually makes the claim that 'no melodic expression associated to a p-licensed position can be more complex than the expression associated to its p-licenser'. This claim is clearly too strong. As I have shown in more detail elsewhere (Brockhaus 1993), it is faced with a vast number of counterexamples, involving onset-nucleus licensing relations. For example, in the German adjective kahl ([kail], 'bald'), the head of the domain is occupied by an expression with a

174

be able to accommodate a more limited range of segmental material than x,. This is exactly what wefind,as the expression associated with x5 contains only a ROOT node and a RESONANCE node (that is, has a complexity of two), while x t has its full complement of class nodes (ROOT, RESONANCE and LARYNGEAL) as well as the elements H, h and ? (giving it a complexity of six). Virtually no other segmental material could be licensed in x5, due to its severely depleted a-licensing potential, while any number of different melodic configurations in x, would still yield possible words of English. Before moving on to the question of how Licensing Inheritance can be related to FOD I need to say a few words about the extremely simple algorithm for determining the a-licensing potential of a position that I have just described, namely straightforward counting of arrows. Although this strategy gives a very good indication of the a-licensing potential available to a position, it is not completely accurate to the extent that there are cases where, for example, it fails to fully predict the drastically reduced a-licensing potential of complement positions within branching constituents. This indicates that the notion of Licensing Inheritance needs to be further refined, a task which would be beyond the scope of the present book. However, as we shall see in what follows, it seems that such an enterprise would be well worth tackling in future work, since a broad range of dynamic processes and static distributional restrictions can be accounted for in terms of Licensing Inheritance. It is perhaps also worth making clear at this stage that depletion of a-licensing potential only occurs to the same extent where the arrows which point to a particular position are in sequence, that is, the first arrow pointing to a position dominates the second and so on. This is the case for all the arrows pointing to, say, x5 or any other position in (21), but it is not true of coda positions which are licensed both by the nucleus to the left and the onset to the right. Here, the two arrows pointing down are not in sequence. As illustrated in (22), which shows the licensing relations in the surname Hendry, the arrow representing the p-licence granted by the nucleus does not dominate the arrow coming from the onset. (22)

[x,]0

[x2

LI X 3 ]N

[X4

xs]0

[XJN

I

I

I

I

I

I

h

ε

η

d

j

ι

Rather than doubly depleting the a-licensing potential of the coda position (x3), the two p-licensing relations represented by these arrows could be thought of as coming close to complexity of only three, while the onset it licenses has a complexity of seven. This, however, does not detract from the basic usefulness of the notion of Licensing Inheritance.

175

cancelling each other out and thus keeping the a-licensing potential of the coda position higher than it would otherwise have been. As suggested in 3.2.1, it is for this reason that the coda in (22) has greater a-licensing potential than the coda (x4) in (23). The representation in (23), incidentally, is that of the English noun laundry. (23)

[xJO

I

Γ

[ [x2

x 3 ] N X 4 ]R [XS

\/

1

X J O [X 7 ]N

I I

3

η

I I

d

J

ι

Returning now to FOD, we need to see what Licensing Inheritance has to say about onset positions which precede a domain-final empty nuclear position. If it turns out that such onset positions invariably exhibit markedly depleted a-licensing potential, then we would predict that the range of segmental material, that is, the degree of complexity, they can accommodate is reduced. To see that this is precisely what we find, consider (24), which shows the licensing relations in brav. (24) [bra:f]

[x, I b

Final Empty Nucleus Parameter ON

\

x 2 ]0 I r

[x3

X4]N

a

[x5]0 I U

NN

ι Studying the arrows, which represent p-licensing, again, we observe the following. The head of the domain, x3, is unlicensed, so that no arrow points down to it. Three positions, χ,, x4 and are directly p-licensed by x3, so that at least X! and x4 are just one step down from the head. The parametrically licensed domain-final empty nucleus which, like x, and x4 is directly p-licensed by x3, is a special case. Although it is at only one remove from the head, x3, it receives additional p-licensing from the Final Empty Nucleus Parameter (7) being set at ON, which is indicated by the second arrow pointing down to x^. In other

176

words, the a-licensing potential of x¿ is doubly depleted. Given that the onset with which the FOD obstruent is associated, x5, inherits its a-licensing potential from it is clear that x5 has severely depleted a-licensing potential. After all, it is at three removes from the head. The crucial question is now whether Licensing Inheritance predicts that there should be a difference in the a-licensing potential of the onset position associated with the alternating labio-dental fricative between a form such as brav ([bra:f]), where FOD applies, and a form such as brave (['bra:va]), where FOD does not apply. To see that this is indeed what Licensing Inheritance predicts, compare the representation of brav in (24) with that of brave in (25).

(25) [braiva]

[xi

I

b

Final Empty Nucleus Parameter ON

x 2 ]0

I

r

[xj

X 4 ]N

[xJO

[xJN

I

I

U

9

1 In (25) the Final Empty Nucleus Parameter (7) is still set at ON, but this setting has no effect, since the find nucleus is filled. Therefore, there is no additional p-licensing of position Xfc which means that this position is less far down the licensing patii in (25) than it is in (24). As a result, its licensing potential is less diluted, so that x5 can inherit sufficient a-licensing potential for accommodating L. Harris (1992) provides evidence from a number of weakening processes in a variety of languages to support his Licensing Inheritance proposals, including Spanish (j-aspiration, lateral depalatalisation, nasal depalatalisation, liquid gliding), Brazilian Portuguese (/-vocalisation to w), Serbo-Croat (also /-vocalisation to w), English (rvocalisation/loss) and Austrian German (/-vocalisation to j). To these I would like to add Modem Greek (Pagoni 1993: chapter 6), where an L spreading from a nasal into an onset with depleted a-licensing potential fails to get licensed, and (Middle) Korean (see Heo 1990), where laryngeal elements remain unlicensed in the same environment. It is important to note that, although the Licensing Inheritance account formally motivates the occurrence of various weakening processes (or distributional restrictions), in particular environments, it has nothing to say about whether a language will actually

177

choose to act on depletion of a-licensing potential. This, of course, is a matter controlled by factors outside the phonology (probably social and historical). For example, languages which lack phonological domain-final reduction phenomena, such as, say, French9 exhibit the same licensing relations as those described for German, but they opt for tolerating depleted a-licensing potential rather than responding to it by segmental reduction. One may speculate that their licensing thresholds are lower, so that even relatively depleted a-licensing potential is sufficient for the licensing of the full gamut of elements. As observed by John Harris (p.c.), all the phonology can do is to make predictions about the range of variation to be expected between languages and to state where phonological processes will occur (as well as identifying universal phonological properties which are common to all languages, of course).10 Another aspect of the issue of variation between languages (and individual dialects of the same language, for that matter) is the question of which elements/class nodes will be affected by phonological processes, or, to put it differently, what kind of weakening processes will occur. Harris's work on /-lenition in English (as reported in Harris 1990, 1994a and Harris & Kaye 1990) makes this point very clearly. Under the same a-licensing conditions, different dialects of English choose to withdraw the a-licence from different elements. In the dialect spoken in New York City, for example, h and ? are unlicensed (resulting in tapping), while in London h and the RESONANCE node suffer this fate (resulting in glottalling). Unlike German, however, English does not appear to exhibit any dynamic weakening processes which affect the LARYNGEAL node. Still, the laryngeal elements seem to be common targets in a variety of languages, which accounts for the fact that, as KLV 1990 assume, there may be no languages where the coda supports contrastive voicing. However, it is not necessarily both laryngeal elements which remain unlicensed. In Modern Greek, for example, only L seems to be affected in this way (see Pagoni 1993). I would argue that this is also true of German, but that, under certain circumstances, not only L but also ? and A fail to be granted a licence. This will be discussed in later sections of the present chapter (especially 4.5.3) and in 5.4.2. We have seen how weakening processes, including FOD, can be motivated through Licensing Inheritance but we have not yet worked out exactly how the problems raised in the context of the discussion of schreiben (see (14), (15) and (16)) can be solved. Recall that my earlier proposal, to delink L, turned out to be problematic because a delinked L would apparently have to be reattached in the course of the derivation of words such as schreiben - a strategy which one would prefer to avoid. It was this 9

10

Liaison, a possible counterexample to the claim that French has no domain-final reduction phenomena, can be dealt with by reference to floating segments (see, for example, Prunet 1986, Charette 1991 (pp. 126ff.) and Durand 1986a), so that no reduction need be involved. Licensing Inheritance also provides implicational hierarchies of environments in which certain processes will occur. In the past, this kind of information had to be captured by means of taxonomic statements such as Escure's (1977) environmental hierarchy discussed in 3.4.3.2.1, whereas Licensing Inheritance does not require such separate scales. All that is, in principle at least, necessary is to consider the licensing paths present in a domain. Weakening processes become progressively unlikely as one moves closer to the head of the domain.

178

particular problem which led to the discussion of autosegmental licensing in the first place. I now propose to interpret the denial of a-licensing in the following way. An element (or node) which cannot be licensed behaves as if it were absent altogether. It is neither phonetically interpreted nor does it make its presence felt from a phonological point of view. For example, it cannot contribute to the complexity value of the expression with which it is associated. At the same time, however, it remains in the representation, so that it can be interpreted should its licensing status change. (This is what Mascaré (1987a) would analyse in terms of delinking and subsequent reassociation).11 No actual delinking is involved, though, a view which is in line with Goldsmith's, who states that 'autosegmental licensing is distinct from association' (Goldsmith 1989: 149). In other words, autosegmental licensing is purely interpretive. Interpretive phenomena do not involve a change in a representation (which would have to be undone again in a Duke of York derivation), so analytic vowel-initial suffixes are no longer a problem. The fact that we are now assuming that FOD is a matter of interpreting the same phonological structure, specifically, the same segmental representation, in different ways, means that the question of a Duke of York derivation does not arise, as no strategy of the type A -*• Β A is employed. What we see throughout the derivation is A only, although it may receive different interpretations, depending on the conditions surrounding it, to be precise, the a-licensing potential it inherits. A further advantage of this approach is that all the arguments put forward in chapter 3 still hold. When I first proposed this account of FOD in Brockhaus 1992b, it raised the issue of whether delinking should be countenanced alongside withdrawal of a-licensing. At the time (op. cit. : 207f), I speculated that reanalysis of events previously dealt with in terms of delinking (e.g. i-lenition in English; see Harris 1990, Harris & Kaye 1990) should be possible. It has now been shown by Harris (1992, 1994a) that this kind of reanalysis can indeed be carried out, so that it seems reasonably safe to conclude that delinking is a device which is no longer required. Saying that GP countenances only composition and decomposition, as I did in 3.4.2, is then accurate only to the extent that decomposition is understood as denial of a-licensing and not as delinking in the traditional sense. Let me now illustrate what happens to an FOD obstruent in a word with a vowel-initial analytic suffix, using the same example as before (schreib-eri), but with a full segmental representation of die alternating obstruent. An element which is present, but not licensed, is shown in brackets.

11

The fact that the element is still present also means that it is available for spreading to an adjacent position.

179

(26) a. 0

Ν

/\ [[χ I 1 J

X I 1 r

/\ χ I 1 a

X 1 1 ι

0 1 X I 1 U I ?1 I 1 h I 1 (L)

Ν

Ν

I χ]

I

0 1 X I

Ν 1 χ Ι I s

χ Ι I 9

Ν χ]

b. 0 r

/ \

[χ I 1 ;

Ν X I 1 r

/ \X

χ I 1 a

1 1 ι

1

U I ?1 I

O i χ Ι I η

Ν I χ]

1

h

In (26a), i.e. on the first cycle, the FOD obstruent is followed by a licensed empty nucleus. As just proposed, such a position is too far down the licensing path for it to pass on sufficient a-licensing potential to the onset occupied by the alternating obstruent. As a result, L is unlicensed. So, if what is contained in the inner set of brackets in (26a) were to be pronounced, we would hear [Jraip].12 However, the derivation continues and the final empty nuclear position is deleted under reduction/the OCP once the inner brackets have been erased. On the second cycle, in (26b), the FOD segment is now adjacent to a filled nuclear position. As described in the discussion of brav and brave (see (24) and (25)), the Final Empty Nucleus Parameter is, of course, still set at ON, but, as the nuclear position following the FOD obstruent is filled, this parameter setting has no effect, that is, the nucleus receives no additional p-licensing from it. As a result, this position is situated higher on the licensing path than the corresponding position in (26a). To be precise, this position is at only one remove from the head of the domain on the second 12

This, incidentally, is exactly what we want here, as the form schreib' (1st sg. pres. ind. with apocopated schwa) is realised as [Jraip],

180

cycle, while it was at two removes on the first. This means that it has greater a-licensing potential, so that the L in the preceding onset can be licensed. The output of the derivation, therefore, contains a voiced [b] in ['/raiban]. Before moving on to a discussion of Vennemann's variable items, I would first like to investigate those second terms in compounds which are vowel-initial in the orthography, but which still permit FOD to apply to a preceding obstruent. Consider the adjectival compound farbecht from (11). Its representation is shown in (27). (27)

R O

R

N\

O N

I I I χ χ X x] I I I I f

[[X

a

Β

ρ

0

N\

O N

1 I\ I I χ χ χ χ] ] I I çI It ? ε

[χ

On the first cycle, the element L in the final bilabial stop offarb- cannot be licensed. In contrast with the derivation of schreiben in (26), later cycles do not provide a licenser for this element either. This is because the second domain is not actually vowel-initial phonologically, as the left-most onset position is not empty. It is occupied by a glottal stop. So, the OCP effect which eliminated the licensed final empty nucleus in schreib- is blocked. The empty nucleus remains adjacent to the labial stop in farb- and, consequently, L remains unlicensed at all levels of derivation. This, of course, does not just hold for compounds, but also for word-final FOD, in those cases where the relevant word is not utterance-final. For the purposes of the representation in (27) I have simply assumed that a glottal stop is present. This is reasonably uncontroversial, as far as the actual realisation is concerned, but it raises the question of what the role of glottal stop is in the derivation. Do we need to assume that this glottal stop is lexically present? If not, what is its source? Let me begin with the first question. If we can show that the presence of glottal stop contrasts with its absence, then we have a good argument in favour of positing lexical glottal stop. However, such a contrast is non-existent, that is, whether the adjective echt is realised as [îfeçt] or [eçt] makes no difference to how it is interpreted by the hearer. It is for this reason that most phonologists (e.g. Vennemann (1982a), Giegerich (1989), Hall (1992b: 58f.) and Yu (1992b: 84-88)) assume that there is no lexical glottal stop. Instead, several of them (e.g. Hall (1992b) and Yu (1992b)) argue for a postlexical epenthesis rule which inserts glottal stop before a foot-initial vowel. Although there is some disagreement about the stage of the derivation at which glottal stop epenthesis is to occur (Giegerich 1989, for example, argues that the rule applies as early as level 1, immediately after syllabification), the assumption most prevalent in recent studies (e.g. Giegerich 1989, Féry 1991, Hall 1992b, Yu 1992b) is that the foot is the correct domain for this rule. Only foot-initial (prevocalic) glottal stop insertion captures the fact that a

181

'firm onset' (which need not necessarily amount to a full-blown glottal stop)13 can be observed both at the left edge of a morphological domain14 and morpheme-internally as a hiatus-breaker before a stressed vowel. The noun Diät ([di'?e:t], 'diet'), whose representation is given in (28), illustrates the word-internal case. (28)

Ο Ν Ο

Ν

O

ι ι ι /Λ I I I I \/ I

χ

d

χ

i

χ

?

χ

χ

e

χ

t

I can see no reason to disagree with this assessment, to which I would like to add that foot-initial position is a strong environment, that is, one in which weakening events are unlikely or unattested (see, for example, Escure's environmental hierarchy in 3.4.3.1 and the various accounts of lenition events in Harris 1990, 1992, 1994a and Harris & Kaye 1990). Why this should be so is clear under the Licensing Inheritance approach. The relevant onset position receives its a-licensing directly from the head of the domain (the stressed nucleus) and so has well-nigh maximal a-licensing potential at its disposal. I would not only agree with the view put forward by other researchers that glottal stop insertion occurs foot-initially (before a vowel), but also with their claims about the noncontrastiveness of this segment. If it is non-contrastive and entirely predictable, it would make sense not to include glottal stop in lexical representation, although I do not subscribe to the view that no redundancy whatsoever should be tolerated in the lexicon. I propose, instead, to handle glottal stop by means of a parameter setting which specifies German (along with unrelated languages such Wolof,15 for example) as a language which cannot incorporate positionless onsets (i.e. onsets which do not dominate a skeletal position) into its licensing hierarchy. This could be phrased as in (29). (29)

Positionless Onset Parameter Positionless onsets can be licensed: ON/[OFF]

The setting of (29) would be ON for languages such as French and English, whereas languages with obligatory epenthetic prevocalic segments (usually [?]), such as German and Wolof, would have it set at OFF. I envisage its functioning as follows. As licensing relations are built within a domain, each position is incorporated into a licensing hierarchy. Positionless onsets cannot be incorporated, so an ambient ?16 is attached to such an onset, which, as I shall explain presently, results in the creation of a skeletal

13 14 15 16

See Giegerich 1989 and Krech 1968 for discussion. That is, next to a left bracket in configurations such as [A], [[A]B], [[A][B]] and so on. See, for example, Dialo 1981: 40ff. and Dialo 1983: 20ff. Recall, though, that ambient elements are essentially in conflict with the GP principle of nonarbitrariness (see 3.2.2 for discussion).

182

position. Note that I do not propose to assume that the empty onsets which eventually end up being realised as firm onsets have an underlying skeletal position. That would amount to letting in lexical glottal stop through the back door. Instead, I argue that onset positions without a skeletal point can be present freely before licensing relations are established. However, once this occurs, these positions must be dealt with in some way. Either they are eliminated through the application of reduction/the OCP or they receive segmental content, provided by the ambient ?. The derivation of the adjective echt would then look something like (30), where the three stages are not meant to represent separate cycles. The idea is that an ambient ? is available to attach itself to any appropriate constituent node anywhere in a lexical representation. (Its positioning near the left edge of the domain in (30) is purely a matter of expositional convenience.) As licensing relations are built, the empty onset cannot be accommodated. Consequently, the ambient ? attaches itself directly to the onset constituent node, as shown in (30b). This strategy may seem somewhat surprising, but, according to Kaye (1988) it was first proposed and motivated by Vergnaud in his 1982 GLOW presentation. Vergnaud argued that such an association leads to the creation of a skeletal point.17 This is shown in (30c), which completes the derivation of echt. (30)

a.

Ν

? b.

χ

χ

χ

ε

ç

t

R

? c.

Ν χ]

χ

χ

χ

ε

ç

t

R O

17

Ο

Ο

Ν χ]

[χ

χ

χ

χ

?

ε

ç

t

Note that no conflict with the Projection Principle arises through the creation of this skeletal point, since no existing licensing relations are undone.

183

This is unproblematic for forms such as echt or Diät, but consider what happens when we attempt to derive a morphologically complex form where the stem is vowel-final and the (analytic) suffix begins with a vowel. The adjective frei with a fem. sg. suffix -e is an example of such a configuration, and a derivation offreie along the lines of that in (30) is presented in (31), where I adopt the practice of not showing an ambient glottal stop, so that it becomes visible only once it is attached to a node. I shall adhere to this practice throughout the book. (31) a.

N

N

/\ X] I rI aI ιI f

[[χ

b.

X

I X] Is

X

Ο

Ν

Ν

/ \ X X/ \ x ] I I I I

Ix ] I

[ [χ f c. *

r

a

ι

O

/\

[ χ ι 1 f

N X 1 1 r

0

/\

[ X I 1 f

/\

X I 1 a

Ν X. 1 1 Γ

X 1 1 ι

a

O

N

? O

/ \ 1 X X X 1 1 1 1 1 1 a ι ?

As indicated by the asterisks in (31c) and (3 Id), the output *['frai?9] is incorrect. The word is actually pronounced ['fraia]. As observed by the authors referred to above, glottal stop insertion does not take place foot-internally, hence the ungrammatically of *['frai?a]. What is it that distinguishes foot-internal contexts from foot-initial ones? Consider the licensing relations in echt and freie, as illustrated in (32a) and (32b) respectively.

184

(32)

a.

[x,]0 I ?

Final Empty Nucleus Parameter ON

[[xJN xJR I I ε

ç

[x4]0 I

[X 5 ]N

t

b. *

[χ, I f

x 2 ]0 I r

[χ, I a

X4]N

I ι

[X 5 ]0

?

I

[X /// shift in preconsonantal position which began in the 13th century (see Szulc 1987: 130). If [z/sl] were indeed branching onsets, then the prediction would be that these clusters should occur word-initially as well as word-internally. As already mentioned, Vennemann's Law of Initials (see 2.4.2.1.2.3 for discussion) applies in GP to the extent that branching onsets can occur freely wherever the relevant licensing requirements are met. In German, this is the case at the beginning of any non-prefixed word. So, this is where we would expect to find [zl] and [si], if they were branching onsets. In the light of these facts, I propose to reject the hypothesis that [b/pn], [g/kn] and [z/sl] form branching onsets. Excluding all of these clusters from the inventory of branching onsets is a move which receives a certain amount of support also from both Rubach 1990 and Vennemann 1968 (see chapter 2 for discussion of the two works). This leaves [b/pl] and [g/kl] to be dealt with. Their wide distribution and their falling complexity slope suggest that they are well-formed branching onsets. I would certainly argue that there are branching onsets which consist of [bl], [pi], [gl] and [kl] clusters. However, it is not clear whether the sequences illustrated in (34) ought to be equated with these clusters. One piece of evidence militating against such an interpretation is the fact

194

that [kl], as illustrated in (34), can be realised as [çl] or [xl]. These fricative + lateral clusters are completely unattested in clear-cut onset positions, so that the [kl] cluster which can give rise to them seems to be different from the genuine branching onset cluster. In other words, [kl] clusters are ambiguous in German. They may constitute either branching onsets or sequences of onsets separated by an empty nucleus. There is only one way to resolve this question and that is to see whether a vowel can be realised between the members of these apparent onset clusters. If they are genuine branching onsets, then no such vowel could ever surface, as no nuclear position is present. If they have the structure proposed for Rodler in (35), then a vowel can be expected to surface. The relevant evidence, not just for [b/pl] and [g/kl] but also for [g/kn], [b/pn] and [z/sl] clusters is shown in (42). (42)

Bibel Zwiebel Kübel übel

fbiibsl] ['tsviibal] [ley ¡bel] ['?y:bal]

'Bible' 'onion' 'vat' 'bad'

bügeln regeln Kugel schmuggeln

[•by :gsln] ['re ¡gain] [kuigsl] ['Jmugeln]

'to iron' 'to regulate' 'ball, globe' 'to smuggle'

eigen Regen Segen

[?aigan] ['reigan] ['zeigen]

'own' (adj.) 'rain' 'blessing'

eben

[?e:ban]

'level' (adj.)

Dusel

['du:zdl]

'dizziness'

The claim that [g/kn], [pn] and [si] do not form branching onsets can, of course, only be maintained if an alternative analysis is available for those words which contain an initial sequence of this type. Some examples of such words are given in (43). Knie kneten Knöchel Knappsack Knoten Knute Knauf

[kni:] [kneitan] ['knύal] fknapzak] ['knoitan] ['knu:te] [knauf]

"knee' 'knead' 'knuckle' loiapsack' Toiot' 'knout' •knob'

195

Gnom gnostisch Gnu Gneis Gnade

[gnoim] ['gnostij] [gnu:] ['gnais] ['gna:dd]

'gnome' 'gnostic' 'gnu, wildebeest' 'gneiss' 'grace'

Phnom Penh Pneumatik

[pnom'pen] [pnoi'ma:tik]

'Phnom Penh' 'pneumatics'

Slum Slawe Slice Slalom

[slam] ['sla:v9] [sia is] ['sla:bm]

'slum* 'Slav' 'slice' (sport) 'slalom'

In an earlier analysis (Brockhaus 1990), I argued that [g/kn] and [pn] sequences are codaonset clusters. In the light of what has been said in the preceding paragraphs, we can only conclude that this approach is no longer tenable. I shall provide a detailed discussion of how to deal with items such as those in (43) in 4.4.2.4. In the meantime, let me summarise what has been said in this section. We have considered two possible alternatives to treating [d/tl], [d/tn], [b/pl], [g/kl], [g/k(ç/x)n], [b/pn] and [z/sl] sequences as containing an empty nuclear position between the obstruent and the sonorant (as illustrated in (35)). These are to interpret the relevant sequences as coda-onset clusters (as shown in (36)) or as branching onsets (as in (38)). We have come to the conclusion that some clusters (those beginning with a voiced obstruent) are definitely not coda-onset clusters (due to the ban on voiced segments in codas), while others (e.g. [pn]) are ruled out by complexity considerations and yet others (e.g. [si]) constitute possible, but perhaps not particularly preferred coda-onset clusters. Looking at the second alternative, we have found that there are severe restrictions on branching onsets, only some of which can be derived from general principles of the theory. We were left with the possibility of analysing [g/kl] and [b/pl] as possible branching onsets. However, both of these alternatives were eventually ruled out when we considered underived forms of the stems involved in the list of Vennemann's variable items (34). All of them exhibited schwa/vowel alternations where they would be predicted only if these sequences are treated as bogus clusters which are composed of two separate onsets with an intervening nuclear position. It is precisely this approach which I intend to adopt to deal with the items in (43) as well as those in (34). Resorting to invoking such inaudible nuclear positions may seem a somewhat suspect strategy, especially because it may look as if such nuclei can be used freely as a kind of stopgap which is kept handy for occasions when a hole appears in the analysis. This is not the case, though. Licensed empty nuclear positions occur only under very closely defined conditions, which will be discussed in the next section.

196

4.4.2.2 Proper government and the Empty Category Principle The conditions under which licensed empty nuclear positions may occur domaininternally are captured in the Empty Category Principle (ECP), which was first proposed in KLV 1990 (p. 219). My version of this principle, which closely follows the spirit of Kaye 1990b (p. 314), is set out below. (44)

Empty Category Principle i. A licensed empty nucleus has no phonetic realisation. ii. An empty nucleus is licensed if (a) it is properly governed or (b) if it is parametrically licensed in languages which license domain-final empty nuclei.30

As already observed, German is a language which parametrically licenses domain-final empty nuclear positions. In German, empty nuclear positions can then automatically remain without phonetic content, provided that they are domain-final. Domain-internal empty nuclear positions, however, receive phonetic content, unless they are properly governed. Proper government is defined by Kaye (1990b: 313) as follows. (45)

Proper Government A nuclear position α properly governs a nuclear position β iff (a) α is adjacent to β on its projection, (b) α is not itself licensed, and (c) no governing domain separates α from β.

If my claim that German [b/pl], [g/kl], [g/kn], [z/sl] and [b/pn] sequences, as well as the [d/tl], [d/tn] sequences which were discussed earlier, contain a licensed empty nuclear position is correct, then the conditions on proper government stated in (45) must be met in all cases. Before this can be checked, it is first of all necessary to establish the direction of proper government in German. Proper government is a form of government at some level of nuclear projection and as such is parametrically variable in its directionality. According to Kaye (1990b: endnote 21), proper government proceeds from right to left in all known cases. So, in the absence of clear evidence to the contrary, I will assume that German conforms to the general pattern and has right-to-left proper government. We can now return to the issue of whether the licensed empty nuclear positions posited for words such as those in (34) meet the requirements for proper government as specified in (45). Consider the representation of Rodler again. It was first introduced as (35) and is reproduced here as (46), with the level of nuclear projection where the proper 30

This version of the Empty Category Principle was proposed before the introduction of the notion of 'magic licensing' (see 3.2.1), which is why 'magic licensing' is not mentioned. 'Magic licensing1, however, is not relevant to the configurations to be discussed here, so there is no need to update the Empty Category Principle to include it.

197

governing relation is established between the two nuclear positions added (second cycle only). (46)

R O

N

O

I χ/ \ χ Iχ I \/ I o

R

I I Ο Ν Ο Ν I I I I I X X X X X I I I Ν

χ r

«-

d/t

1

a

r

The arrow indicates the proper governing relation which holds between the penultimate nuclear position and the empty nucleus which precedes it. All the relevant requirements are met, as the governor is not itself licensed (hence its phonetic content), the two are adjacent at the relevant level of nuclear projection and there is no governing domain intervening between them. A glance at (34) will confirm that this is true also for all the remaining items.31 It appears, then, that for NSG speakers it is not just parametrically licensed domainfinal empty nuclear positions which are unable to license L in a preceding onset (and which thus trigger FOD), but also properly governed domain-internal empty nuclei. Speakers of Hochlautung, on the other hand, distinguish between the two types of licensed empty nuclei. For them, it is only parametrically licensed domain-final empty nuclear positions which cannot license L. The fact that the two behave in non-identical ways is by no means an idiosyncratic property of German. The same kind of distinction between the two types of licensed empty nuclei has been observed for Modern Greek (see Pagoni 1993: chapter 6 and Brockhaus 1992a for discussion). However, it is not just with regard to autosegmental licensing that properly governed empty nuclei are distinguished from parametrically licensed empty nuclei. Charette (1990, 1991: 129ff., 1992) discusses the notion of government-licensing, which she invokes to account for the absence of branching onsets and coda-onset clusters preceding licensed empty nuclei which receive no phonetic interpretation. In her view, an onset head can govern a complement within a branching constituent or a coda only if it is given licence to govern by its nuclear licenser. Government-licensing predicts that a licensed empty nucleus either receives phonetic content32 in order to be able to discharge its government-licensing responsibilities or that a possible governing domain preceding a licensed empty nucleus without phonetic interpretation is broken up, because the

31

32

What the proper government account put forward here does not deal with is the fact that the FOD obstruent in Vennemann's variable items is always followed by a sonorant. This is not what one would expect in the context of proper governing relations. Instead, there should be no restrictions on the segment types surrounding the properly governed empty nucleus. I shall return to this issue in 4.4.2.5. That is, it behaves as though it were unlicensed in the sense of the Empty Category Principle (44).

198

governing onset cannot be granted a licence to govern. In Charette 1992, she examines several languages and comes to the conclusion that at least in one dialect of French and in Khalkha Mongolian (Altaic; Mongolian People's Republic, China) domain-final empty nuclear positions and properly governed empty nuclei differ in their ability to government-license a preceding onset head. It is perhaps striking, though, that, where such differences can be observed, it is usually the domain-final empty nuclear position which has greater government-licensing potential than the properly governed empty nucleus. This is the exact opposite of what we find for autosegmental licensing. It is too early to say whether this is in any way significant, which is why I shall not discuss it further. Perhaps it means nothing more than that government-licensing and autosegmental licensing potential are independent of each other.

4.4.2.3 An alternative solution: differences in morphological structure between Hochlautung and NSG There is a possible alternative to the interpretation that Vennemann's variable items contain a domain-internal licensed empty nucleus, which may well be worth considering. Suppose that for NSG speakers words such as those in (34) have a different morphological structure from the one they have for speakers of Hochlautung. In particular, they may have the agent suffixes -1er (e.g. Rodler and Pendler) and -ner (e.g. Eigner and Ordner) instead of the standard -er. This assumption would be supported by productive use of the suffixes -1er and -ner. Before considering evidence from productivity in more detail, let me first provide some examples of nouns derived by means of these suffixes (transcriptions reflecting Hochlautung). (47)

a.

Künstler Dörfler Sportler Wissenschaftler

['kYnstle] ['dœçfle] ['/poçtlB] ['visanjaftte]

'artist' 'villager' 'athlete' 'scientist'

Ländler Häusler Nachzügler

['lentie] ['ho isle] ['na:xtsy:kte]

'country waltz' 'cottager' 'straggler'

Glöckner Rentner Schaffner

['glceknu] [rentne] [Jafnu]

"bell-ringer' 'pensioner' 'conductor' (transport)

Redner Söldner Schuldner Bildner

[re:dne] f'zœldnB] ['Juldntj] ['bildnu]

'speaker' 'mercenary' 'debtor' 'sculptor'

199

If the difference in FOD behaviour between the two groups of speakers is indeed due to the fact that NSG speakers have domain-boundaries after every obstruent to which they apply FOD, then the prediction would be that only these speakers use the suffixes -1er and -«er productively. However, Fleischer & Barz (1992: 156ff.), who - as far as one can ascertain - describe the standard variety, observe that both suffixes are used productively, although -1er is more productive than -ner. In other words, the productive use of -1er and -ner is not restricted to NSG. There is, however, an interesting difference between NSG and Hochlautung with regard to -ner which is brought to light by the data in (47). The Hochlautung pronunciations shown are identical to the corresponding NSG pronunciations in (47a), but not in (47b). In (47b, second group), FOD applies for NSG speakers, but not (according to Mangold et al. 1990) in Hochlautung. Assuming that speakers of Hochlautung apply FOD before licensed domain-final empty nuclei, but not before domain-internal licensed empty nuclei, the only conclusion one can draw from this is that for these speakers, the -n must be part of the stem (presumably a kind of complex stem consisting of the root and a non-analytic suffix -(e)ri). Only if that is the case can the empty nucleus which precedes it be treated as domain-internal. In other words, these speakers interpret the relevant words as having stem-final -n followed by the familiar agentive suffix -er. So, although Hochlautung speakers recognise and use the suffix -1er, the suffix -ner does not appear to exist for them, hence the very limited productivity. Apart from the limited productivity, there is a noticeable amount of variation in the FOD behaviour of words such as those in the second group of (47b), which also suggests that the status of -ner is uncertain for Α-speakers. In my informant work, Bildner, for example, was pronounced both ['biltne] and ['bildne] by the same speaker in a single session. To illustrate my point about the differences between -1er and -ner for speakers of Hochlautung, I have added morphological bracketing and formative boundaries (which are invisible to the phonology, as already noted) to the data from (47) in (48). The morphological bracketing ([ ]) is to be understood as described in 3.2.2. That is, square brackets identify an analytic morphological domain. (This means that the brackets correspond roughly to SPE's (pp. 366ff.) # and ##.) The formative boundary (+) is present in non-analytic morphology (as defined in 3.2.2), but it is not visible to the phonology. It thus has quite a different status from the +-boundary referred to in SPE (pp. 364ff.), which, by contrast, is visible to the phonology. (48)

a.

[[Künstjler] [[Dörf]ler] [[Sport]ler] [[Wissenschaftler] [[Länd]ler] [[Häusjler] [[Nachziig]ler]

['kvnstle] ['dœçfhj] ['Jpoçth?] [Visan/aftle] ['lenti«] ['ho isle] ['na:xtsy:kle]

'artist' 'villager' 'athlete' 'scientist' 'country waltz' 'cottager' 'straggler'

200

b.

[[Glöck+n]er] [[Rent+n]er] [[Red+n]er] [[Söld+n]er] [[Schuld+n]er] [[Bild+n]er]

['glœknB] ['rentne] ['reidriß] ['zœldne] ['Juldnu] [•bildnB]

"bell-ringer' 'pensioner' 'speaker' 'mercenary' 'debtor' 'sculptor'

What emerges from this discussion is that the morphological structure of certain words is not always as straightforward as it may at first appear. It seems that Hochlautung speakers work on the assumption that the words in (48b) are composed of a complex stem ending in -n followed by the suffix -er, instead of a simple stem (consisting only of the root) followed by -ner. Interestingly, all the words involved in this have common forms ending in -n in their inflectional paradigms, usually in the nominative plural for nouns (e.g. Glocken "bells') or the infinitive for verbs (e.g. bilden 'to form'). It is, perhaps, this sort of suffix which provides the nasal at the end of the complex stem. The words in (48a), which contain the suffix -1er, by contrast, invariably lack the lateral in these inflected forms (e.g. Künste 'arts', nom. pl.; Dörfer, 'villages', nom. pl.). It is also clear now that both Hochlautung and NSG have the suffix -1er and there appears to be no evidence to suggest that NSG speakers use this suffix differently from Hochlautung speakers. However, as far as -ner is concerned, things are less clear. This suffix does not appear to be present in Hochlautung, but it may be in NSG. Figures on productive use are, to my knowledge, unfortunately, not available, and there is no phonological evidence, since the analysis predicts identical FOD behaviour for both -er (when preceded by [n]) and for -ner in NSG. Returning to Vennemann's variable items (see (34)), though, referring to differences in morphological structure in order to account for the difference in FOD behaviour between NSG speakers and Hochlautung speakers does not seem to work at all well. It is problematic because, as already pointed out in 2.4.2.1.2.1, it requires us to posit stems (such as pend- or eb-) for NSG speakers which cannot be independently motivated. This argument also militates against yet another alternative (also discussed in 2.4.2.1.2.1), which is to posit three morphemes to account for NSG speakers' pronunciations of the variable items. This proposal is made by Kloeke (1982a; see 2.4.2.2) for speakers of Hochlautung, but it would not work for these speakers given my analysis. It could do, though, for NSG speakers. Pendler would then be bracketed as [[[Pend]l]er], The predictions for FOD would be correct, but the root form would still be exceptional, just as it would have been under the assumption that only two morphemes are involved. To sum up, given the GP account as developed so far, the difference in FOD behaviour illustrated in (34) can be attributed to one of two factors. Firstly, it can be due to the fact that the two dialects respond differently to the depletion of a-licensing potential associated with parametrically licensed domain-final empty nuclear positions as opposed to properly governed empty nuclei. Secondly, it may be an effect of a different assignment of morphological structure between the two groups of speakers. I

201

propose to reject the latter, as, among other things, it would predict the existence of sterns such as pend-, eb-, rod-, etc., which cannot be motivated independently. In the remainder of this chapter I will investigate the implications of adopting the former view.

4.4.2.4 More on problematic clusters Let me now return to die problematic word-initial clusters [g/kn], [pn] and [si] illustrated in (43). I proposed to analyse [g/kn] and [pn]33 so that the obstruent was separated from the following sonorant by a licensed empty nuclear position. To see whether this proposal is compatible with the requirements in (45), consider the representation of Gneis in (49). (49)

R

s

w

A A you Ί1 kann ste

The input to Defooting (85) is precisely the prosodie structure which predicts the forms in column I of (84). Taking the example oïhab' ich, Fs' would be built on [ha:p] and Fw2 on [?iç], so that the glottal stop would indeed be foot-initial. The application of (85), however, changes this. The deletion of Fw2 means that its contents now become part of F1. In other words, the glottal stop is no longer foot-initial; it is part of the weak member of a foot. In terms of licensing relations, the onset to which the ? element is attached is not directly licensed by the head of the domain. Instead, it receives its licensing from a nuclear position which is itself licensed by the head of the domain (i.e. by the nuclear head of [ha:p]). As a result, the a-licensing potential this onset inherits is sufficiently reduced for ? to remain unlicensed. In other words, the output of Defooting (85) generates [ha:p iç]. It seems that the application of Defooting (85) causes some existing licensing relations to be undone (by the deletion of a foot node) as well as leading to the addition of new licensing relations. The latter is compatible with the Projection Principle (see 3.2.2), but the former is not. To solve this problem, we may have to say that the forms in column Π of (84) are not derived from the corresponding forms in column I, that is, the two are represented separately in the lexicon. Finally, we need to find appropriate representations for the forms in column III of (84). This should be reasonably straightforward, since they exhibit the same behaviour (with regard to FOD and the presence or otherwise of glottal stop) as single words (which may or may not have internal analytic morphological structure). We can then assume that appropriate lexical representations would be along the lines of that in (86). This involves an analytic structure with two domains rather than three, i.e. [[A]B], which would enable reduction to apply and thus produce a filled nucleus to license an L in a preceding obstruent, while, at the same time, capturing something of the partial independence of ich from the verb. The relevant derivation is shown in (86).

236

a.

[ [

b.

[

0 1 X 1 h 0 1 X 1 h

Ν

/ \χ \/

χ

0

Ν

Ν

I χ

I χ ]

I χ

χ

I I

I ç

a

I Ρ

Ν

0

Ν

0

Ν

I χ I b

I χ I I

I χ I ç

I χ ]

/\ χ \/

χ

a

Ν I χ ]

An alternative to this would be to assume that what may have started out as cliticisation (or affixation) has found its way into the non-analytic morphology of German, so that forms such as /ha:b+iç/ actually constitute separate lexical entries. Be that as it may, the discussion in this section illustrates the point that domainboundaries and (orthographic) word-boundaries do not necessarily coincide. In German, the two do coincide in the vast majority of cases, but this cannot be taken for granted, and the failure of FOD to apply word-finally is an indication of the fact that a different configuration with regard to domain-boundaries is involved. In this case, ich does not occupy a domain of its own.

4.7 Conclusion In this chapter I have investigated the question of where FOD occurs, in the framework of Government Phonology. I have observed that an FOD environment contains an FOD obstruent occupying an onset which inherits its a-licensing potential from a licensed empty nuclear position. In the light of the fact that such nuclei have been shown to have reduced a-licensing potential in a number of languages (e.g. English, Modern Greek, Korean, Brazilian Portuguese, Spanish, Serbo-Croat and Austrian German), I have proposed to reinterpret FOD as an instance of autosegmental licensing. Specifically, I have argued that the laryngeal element L cannot be licensed by positions with depleted a-licensing potential. German is a language which parametrically licenses domain-final empty nuclear positions, and I have shown that these positions cannot license L in a preceding onset,55 so that FOD applies domain-fmally, provided that no filled nucleus becomes available during the derivation.

55

In chapter 5, I will discuss the spirantisation of underlyingly voiced velar stops in FOD environments. It may be that ? is another element which, under partially similar circumstances to those discussed for L (and A), cannot be a-licensed by a licensed empty nucleus.

237

By reanalysing Vennemann's variable items in the GP framework, I have also been able to account for the pronunciation difference between speakers of Hochlautung and speakers of NSG. This is due to the fact that Hochlautung speakers distinguish between parametrically licensed domain-final empty nuclear positions and domain-internal nuclei which are p-licensed through proper government or interonset licensing. For these speakers, only the former are unable to license L. This distinction between the two types of licensed empty nuclei is not unique to German, as it has been described for other languages, e.g. Modem Greek (Pagoni 1993), French and Khalkha Mongolian (Charette 1991: 129ff, 1992). NSG speakers, on the other hand, draw no such distinction between the two types of nuclei, so that FOD applies preceding both. The GP analysis has also enabled me to shed light on some of the problems with Rubach's (1990) account. It has become clearer why speakers' unmonitored 'phonological behaviour" occasionally differs from their intuitions about syllabification. Furthermore, I have made some suggestions about the reasons for the failure of FOD to apply in obstruent + /r/ clusters in NSG. Apparently, this is connected with /r/'s particular propensity to spread into a preceding nucleus, which, in turn, is a direct consequence of the fact that /r/, unlike ñJ and /n/, is composed only of elements which occur in nuclear positions anyway.

5. Final Obstruent Devoicing and Neutralisation

5.1 Introduction For over 50 years, the phonological event of FOD has been cited as one of the prime examples of phonological neutralisation. This view of it, however, has become quite controversial over the past decade and a half. A lively debate about whether it is mistaken or notfillsmuch of the experimental literature of the 1980s (see, for example, Port et al. 1981, O'Dell & Port 1983, Dinnsen & Charles-Luce 1984, Fourakis & Iverson 1984, Charles-Luce 1985, Port & O'Dell 1985, Port & Crawford 1989; also Dinnsen 1985 and Fourakis 1984 for discussion). In this chapter,11 will try to unravel some of the issues intertwined in this debate and see what phonologists can leam from a discussion which involved mainly phoneticians, but which was clearly sparked off by claims made in the phonological literature. I will begin with a very brief overview of parts of the history of the theoretical construct of neutralisation, which will be followed by a discussion of the most important findings of the experimental work on FOD in German in the context of phonological neutralisation. I will point out some of the implications of this research for the phonological theories which inspired this work in the first place, but which do not appear to have benefited from it. I will then provide a brief reminder of the relevant aspects of the GP analysis of FOD, as presented in chapters 3 and 4. From the vantage point of this element-based (rather than binary feature-based) analysis, I will be in a position to identify some problems raised by the experimental studies as being artifacts of the phonological framework the studies took as their point of departure. I will show that some of the basic tenets of GP go at least some way towards making it possible for the remaining problems arisingfromthe experimental work to be resolved.

5.2 Neutralisation and the traditional place of FOD 5.2.1 Two views of neutralisation: Trubetzkoy and Kiparsky According to Davidsen-Nielsen (1978: 26f.), the notion of neutralisation was first introduced by Trubetzkoy in two papers published in 1933 (Trubetzkoy 1933a, b). In his seminal work Grundzüge der Phonologie, published six years later (Trubetzkoy 1939 [1969]), Trubetzkoy uses the term neutralisation to cover a much wider range of possibilities than most writers on the topic after him. This high degree of differentiation 1

An earlier version of this chapter appeared as Brockhaus 1991.

239

is made possible through his use of the concept of the archiphoneme (first introduced by Roman Jakobson in Jakobson 1929), which he defines as 'the sum of distinctive properties that two phonemes have in common' (op. cit. : 79). In his view, there are four different cases of neutralisation to be found in natural languages. In the first case, neither of the opposition members involved in the neutralisation is actually identical with the realisation of the archiphoneme in the position of neutralisation. In other words, given an opposition between the segments a and β, what appears in the neutralisation environment will be a third segment γ, where γ has some properties of both a and β, without being identical with either of them. An example of this type of neutralisation would be the English flapping rule, which merges the contrast between l\l and /d/ into a new segment (often transcribed as [D]) which differs from both the two input segments (see e.g. Dinnsen 1985: 266). In the second case, one of the opposition members is identical with the phonetic realisation of the archiphoneme in the position of neutralisation. The choice of which member occurs here is conditioned 'externally', in Trubetzkoy's terms, i.e. is dependent on the phonological properties of an adjacent phoneme. Voicing assimilation of obstruents (e.g. in Polish or Russian) would be an example of this case. The third case is like the second, except for the fact that the choice of which member of the opposition represents the archiphoneme is conditioned 'internally', that is, in a privative opposition, the unmarked member of the pair will be chosen in the neutralisation environment by virtue of its being unmarked. FOD could serve as an example here, provided, of course, that the realisation of, say, final /d/ is actually identical with the realisation of /t/ in the same environment. The fourth case, which Trubetzkoy concedes is rather rare and very often consists of a combination of the second and third cases, has both members of the neutralised opposition representing the archiphoneme in the neutralisation environment, one in one type of environment and the other in another. He gives the opposition of German [s] [J] as an example of a genuine (i.e. non-combinatory) instance of case IV. He argues that this opposition is neutralised before consonants, with [J] occurring root-initially and [s] root-medially and finally (p. 82). Having seen how differentiated a view of neutralisation was adopted when the concept was first introduced, we can now skip nearly four decades of further development and arrive at the approach to neutralisation which provided the basis for most of the experimental work to be discussed in the remainder of the present chapter. This more recent approach is that put forward in Kiparsky 1976 (p. 169), which is much more restrictive than Trubetzkoy's. Here, the concept of neutralisation plays a crucial role in the phrasing of the Alternation Condition (Kiparsky 1976: 167Í), which was intended to constrain the abstractness of underlying representations by limiting the application of phonological rules so that absolute neutralisation processes were eliminated. The definition of neutralisation in this context takes the form in (1).

240 ( 1)

Suppose we have a phonological process P: (P)A->B/XC_DY where C and D represent a (phonological and/or morphological) context, and X and Y are arbitraiy strings. Then, a.

Ρ is NONAUTOMATic if there are strings of the form CAD in the immediate output of P; o t h e r w i s e Ρ is AUTOMATIC.

b.

Ρ is NEUTRALISING if there are strings of the form CBD in the immediate input of P; o t h e r w i s e Ρ is NON-NEUTRALISING.

(emphasis his)

What this definition says is that, for a process to be neutralising, the output Β which it generates must already be present in the system. If Β (or, more precisely, CBD) is not available as a potential input to P, then the process is non-neutralising. In other words, Trubetzkoy's case I is excluded. So, English flapping, for example, would no longer qualify as a neutralisation process. This definition of neutralisation appears to have been accepted fairly generally in generative phonology, and some analyses crucially depend on it (e.g. Houlihan & Iverson 1979). Kiparsky 1982a, however, marks the beginning of a substantial decline in the importance of neutralisation as a theoretical construct. The Alternation Condition, to which neutralisation was so important, is exposed as being seriously flawed, for, among other things, 'it is ... not a formal condition of the desired sort because the property of being a "neutralisation rule" is not determinable from inspection of the grammar' (Kiparsky 1982a: 152). Except for informal use (which I will briefly touch upon at the end of this chapter), the term 'neutralisation' seems to have all but disappeared from phonological discussion today. Its spirit, though, lives on in the regulative principle of Structure Preservation in Lexical Phonology (see, for example, Pulleyblank 1986 (p. 7), Kaisse & Shaw 1985, Kiparsky 1985). A structure-preserving rule is very similar indeed to a neutralisation rule by Kiparsky's 1976 definition. Structure Preservation may be somewhat less restrictive in that it does not require elements C and D of Kiparsky's 1976 definition of neutralisation. For Β to be underlyingly present is sufficient. Also, neutralisation can be interpreted as a special case of Structure Preservation in that it relates to segmental structure only. Structure Preservation (at least in its more recent formulations; see Borowsky 1989)2 also covers prosodie structure, such as constraints on the number of positions within a coda for example.

2

She observes that the definitions of Structure Preservation advanced by Kiparsky and herself in 1985 and 1986 respectively suggested (without being entirely clear on this) that only segmental constraints were structure preserving. In her paper, she makes a strong case for explicitly including prosodie structure in the domain of Structure Preservation.

241

5.2.2

Earlier analyses of FOD and their implications for FOD as a neutralisation process

In this section, I will briefly review those aspects of earlier analyses of FOD which are relevant to the question of whether FOD constitutes a neutralisation process or not. What we need to focus on is obviously the 'what' of FOD, that is, the issue of what exactly changes when an obstruent undergoes devoicing. As discussed in detail in 2.2, we can distinguish between two basic positions here. Simplifying the various proposals somewhat, we can say that the first is based on binary features and makes the claim that a segment which has underlying [+ voice] undergoes some kind of process which results in that specification being changed to [- voice].3 The second, more recent, by contrast, makes reference to privative rather than binary features and, roughly speaking, equates the [+ voice] specification of earlier work with the presence of the privative feature [voice] and the absence of this feature with what had previously been handled in terms of a [- voice] specification.4 Given both of these analyses, FOD obviously qualifies as a neutralisation process in Trubetzkoy's terms, an instance of case III. That it also meets the requirements of Kiparsky's more restrictive 1976 definition is made clear by the data set in (2). The lefthand column shows words which would constitute 'strings of the form CBD' as input to Ρ and the right-hand column contains words with apparently identical CBD strings (as far as the pronunciation, but not necessarily the morphological structure is concerned) which were changed from CAD to CBD by the application of FOD. The relevant environment is emboldened and italicised in the orthography. (2)

3

4

Alptrmim [?alptraum] 'nightmare'

ha/6^rocken ['halptrokan] 'medium dry'

altbacken [?altbakan] 'old-fashioned'

WaWfteere [Valtbeira] 'fruit of the forest'

Werftwohnung ['veçkvornuq] 'company-owned flat'

Bergwacht [teçkvaxt] 'mountain rescue service'

Scha/Äerde ['Ja:fhe:çda] 'flock of sheep'

Bravheit ['bra:fhait] 'good behaviour'

This approach is proposed, for example, in Vennemann 1968, 1978, Wurzel 1970, 1981b, Hall 1989a, 1992b (p. 53), Rubach 1990 and Wiese 1988 (p. 80), 1991. The most notable proponents of this strategy are Goldsmith (1989,1990) and Lombardi (1991).

242 The earlier analyses of FOD, then, leave no doubt about its neutralising status, regardless of whether one chooses to subscribe to a broad (Trubetzkoy) or a narrow (Kiparsky) interpretation of that concept. The period of increased importance for neutralisation as a theoretical construct which began in the early 1970s and came to a close with Kiparsky 1982a partly overlapped with a time of change in phonetics. There was a marked shift in the attitude of phoneticians towards the relationship of their science with the more theoretical aspects of linguistics, and phonology in particular. Well into the 1970s, the general consensus had been that phonetics was charged with establishing how exactly the linguistically determined categories were borne out in the 'real world' of speech sounds. This view met with increasing dissatisfaction from practising phoneticians, resulting in a conscious effort to establish 'experimental phonetics as a serious contribution to theory and explanation in phonology rather than as a mere field of additional concrete descriptive statements' (Kohler 1984: 151). The title of Ohala's 1974 paper ('phonetic explanation in phonology') constitutes a kind of rallying cry of this new approach to phonetics. Neutralisation was one of the first issues to be tackled, with final devoicing in German, the textbook example of it, attracting most attention (see 5.1 for references).

5.3 The neutralisation debate in the experimental literature 5.3.1 The studies and their findings Most of the work was carried out in the United States. As far as one can tell, all researchers based their work on Kiparsky's 1976 definition of neutralisation, although only a small minority state this explicitly. What is clear, though, from the ways the studies are conducted is that neutralisation is interpreted as a devoiced obstruent becoming indistinguishable from its underlyingly voiceless congener. It turned out that it was extremely difficult to establish this reliably. The most important problems were connected with the set-up of the relevant experiments and the choice of test words (and, occasionally, the voicing status assigned to the underlying segments; e.g. Mascaró 1987b). One of the objections taken most seriously was that experiments which involved reading lists encouraged so-called spelling pronunciations. The problem was that, in a language such as German, the spelling of the alternating segment remained unchanged, regardless of whether it was pronounced as voiced or voiceless (see 1.2.2.2 for examples). It was argued (e.g. in Fourakis & Iverson 1984) that this 'voiced' spelling suggested to the experimental subject a voiced sound or at least some residue of voicing, which then led to a distortion of the results. The researchers explored two different ways of overcoming this problem. Firstly, in studies on German, test words were elicited without reading lists by orally cueing subjects for an infinitive form of a verb which had to be conjugated. The preterite form contained the desired FOD obstruent (see Fourakis & Iverson 1984). Secondly, experimental work was carried

243

out on languages such as Catalan, where the spelling of FOD obstruents generally changes with their voicing value (see Dinnsen & Charles-Luce 1984, Mascaré» 1987b and Charles-Luce & Dinnsen 1987). The findings were less clear-cut than the researchers had hoped. Some studies had negative results, i.e. no statistically significant differences between underlyingly voiceless obstruents and their devoiced congeners were identified (e.g. Fourakis & Iverson 1984 for German and Jassem & Richter 1989 for Polish). Others, however, reported positive results, i.e. statistically significant differences between these groups of voiceless obstruents had been observed (e.g. Port et al. 1981, O'Dell & Port 1983, Port & O'Dell 1985 and Port & Crawford 1989 for German; Slowiaczek & Dinnsen 1985 for Polish and Dinnsen & Charles-Luce 1984 and Charles-Luce & Dinnsen 1987 for Catalan). It is hard to say what exactly the reasons for the differing results were, but it appears that the details of the experimental set-up and the statistical methods of analysis may have been decisive. There can be no doubt that the differences in the production of devoiced obstruents and underlyingly voiceless obstruents are very small indeed, but, in my view, those researchers who observed these differences and showed that they are statistically significant have a better case than those who argued that the differences are non-existent or not significant. Firstly, the former used a much wider range of different experiments and statistical methods than the latter, and still succeeded in consistently coming up with similar findings. Secondly, as Dinnsen (1985) points out, the failure to find statistically significant differences does not necessarily mean that they are not there. Perhaps the researchers were looking in the wrong place. It seems reasonable, then, to conclude that, at least for the languages examined and as far as production is concerned, neutralisation may not be complete. Before discussing the findings and the issues they raised further, let me first describe the general pattern of the experimental design to which most of the studies conformed, to show how these findings were actually arrived at. This section covers studies on German only. Words containing alternating obstruents in FOD environments (i.e. realised as voiceless) were elicited from native speakers together with similar words with underlyingly voiceless obstruents, so that sets of minimal pairs (e.g. Rad 'wheel' and Rat 'advice') or near-minimal pairs were recorded (the focus of the experiment was usually disguised by mixing these words with others not relevant to the study). These recordings were then used for measurements of various temporal aspects of the pronunciations of these words. Most commonly, the researchers measured the duration of the vocalic nucleus preceding die obstruent under investigation, the duration of the period for which glottal pulsing continued from the preceding nucleus into the closure or friction phase of the obstruent itself and the duration of the closure or friction phase of that obstruent. As shown in (3), it was generally found that underlyingly voiced obstruents differed from their underlyingly voiceless congeners in that the preceding vowel was longer, the duration of voicing into closure or friction was greater and the duration of the closure or

244

friction phase was shorter. It was not necessarily the case that each study found statistically significant differences for all three variables at the same time. (3)

Phonetic variables

Most common findings Underlyingly voiced voiceless

Duration of preceding vocalic nucleus Duration of voicing into closure/friction (from the left) Duration of closure/friction

longer longer

shorter shorter

shorter

longer

To see whether the contrast which was maintained in the production of obstruents was also perceptually salient, listening tests were carried out. For this purpose, recordings of a large number of tokens (usually the words used in the production experiments) were played to native speakers of German (not those used in the production experiments, of course), whose task it was to pick out the form they believed they had just heard from two choices presented to them on an answer sheet. The overall percentages for correct responses are shown in (4). In some cases, these are the means of results from several experiments conducted as part of a single study. (4)

Study

Percentage of correctly identified tokens

Port et al. 1981 O'DellÄ Port 1983 Port & O'Dell 1985 Port & Crawford 1989

70% 63% 59% 69%

The relatively wide range of these results is due to differences in the experimental set-up (in one case, for example, speakers were told to dictate the words to someone when the recording was made, so their pronunciation was exaggerated to reflect the spelling interestingly, however, not to such an extent that the listeners were able to identify much more than about 70% of the words correctly). Whatever the most representative figure may be, it is clear that the level of correct responses cannot be due to mere guessing. The performance of the subjects may have been poor, but it was still consistently better than chance. The researchers concluded that the underlying voicing distinction is maintained to the extent of being perceptually salient. However, it is not clear whether it is salient enough to be relied upon in normal communication. Apart from the original issue of neutralisation, to which I shall return shortly, a number of other questions were raised by this experimental research.

245 Most of tfaem fall into one of two categories, the question of the relationship between production and perception, and the issue of how exactly the output of the phonology is implemented phonetically. I shall discuss them in turn, starting with the issue of production vs. perception.

5.3.2 The relationship between production and perception It is clear from (3) and (4) that there is a certain mismatch between production and perception concerning the distinction between underlyingly voiceless vs. devoiced obstruents. The production facts suggest that the two are different, as do the perception facts - but only to a point. Even if 60-70% of all word-final voiceless obstruents can be correctly identified as underlyingly voiceless or devoiced by a native speaker-hearer, that may still leave too great an error margin for anyone to wish to say that this distinction is used to carry information crucial for successful communication. If the perceptual contrast between the two groups of voiceless obstruents, then, is too small to be important, one could argue that, as far as perception is concerned, neutralisation is perhaps complete after all. One conclusion one could draw from this is that the statistically significant differences in production are linguistically not significant. How can this mismatch between production (where consistent differences have been observed) and perception (where the difference is only marginally salient) be captured in a grammar which is neutral with regard to speaker vs. hearer? This neutrality is, after all, one of the basic assumptions about the grammar in current mainstream thinking in generative phonology.5 The question of how to deal with this mismatch is discussed by Dinnsen (1985) as part of his review of the concept of neutralisation. He concludes that it is impossible to reconcile such differences between production and perception, and suggests that production and perception should be seen as being at least partially independent of one another. (See also Klatt 1981 for detailed arguments in favour of the dual-lexicon hypothesis.) Obviously, the dual-lexicon hypothesis is too big an issue for a satisfactory discussion in a book such as the present. What matters for our purposes is that, from the phonologist's point of view, it would be preferable to be able to work with a single lexicon. This would avoid additional problems, such as a substantial complication of the

5

A mismatch between production and perception has also beai reported in the context of historical mergers (see e.g. Harris 198S: 334ff, Nolan 1986). Of the explanations advanced for this I find the view that fine auditory distinctions made by die child are not maintained later in life most interesting (attributed to Drachmen by Linell (1979: 42)). The idea is that children perceive any consistent distinctions present in adult articulation and form articulatory habits of their own which incorporate these distinctions. The ability to discriminate so finely is, however, lost with age, so that adult speakers who produce two distinct segment types are unable to tell them apart. As I am not familiar with conclusive arguments concerning this view, I shall not pursue it further. In any case, Drachman's ideas are compatible with my own proposals (see 5.4.4), so that it is not vital to make a decision here.

246

grammar and questions regarding the interaction between the two lexicons. Ideally, one would hope that a single lexicon could accommodate both production and perception. (See Fay & Cutler 1977 for evidence from production errors which supports the view that there is just one lexicon.) Whatever one's view of other arguments advanced in favour of the dual-lexicon hypothesis may be (e.g. in Klatt 1981), as far as FOD is concerned, it is my contention that the mismatch between production and perception can be resolved without resorting to the dual-lexicon. I will discuss this below, in the context of my GP analysis of FOD.

5.3.3 The implementation of FOD The second additional issue raised by this experimental work concerns the phonetic implementation of FOD. The researchers wondered how the systematic production difference between underlyingly voiced obstruents and their underlyingly voiceless congeners in FOD environments could be reconciled with the prevailing assumption that the two are identical phonologically.6 One way of dealing with this problem is by means of phonetic implementation rules.7 Specifically, it was suggested (in Port & O'Dell 1985) that phonetic implementation rules should not just be understood as implementing segments and segmental features, but as implementing properties of whole syllables as well. In the context of FOD, the idea was that an underlyingly voiced syllable-final obstruent would not undergo a devoicing rule in the phonology, but would become the input to the relevant implementation rule with its [+ voice] specification intact. The syllable implementation rule would then initiate 'a gesture that resembles the segmental implementation of [- voice]' (p. 468). This would guarantee that both [- voice] and syllable-final [+ voice] segments were pronounced in a very similar way, but not identically. Implementation rules, like the dual-lexicon hypothesis, are a complex and widelydebated issue (at least in the phonetic literature), and thus an issue which cannot be done justice to in a book such as the present. I will, therefore, consider only how FOD would fare in a grammar incorporating both phonological rules and phonetic implementation rules. An important question which has been discussed by proponents of such implementation rules is how exactly they interact with phonological rules. One of the clearest recent statements of this relationship I am aware of comes from the phonological (Lexical Phonology) rather than the phonetic literature. Pulleyblank (1986: 7f.) suggests that the phonetic rules (I believe that these are essentially the same as Port & O'Dell's (1985) implementation rules) should apply after all phonological rules, that is, the input 6

See Nolan 1986 for a brief discussion of some possible answers to this question. Phonetic implementation rules, which appear to amount to the same thing as phonetic realisation rules, are, of course, nothing new. They have been discussed in the literature at least since the days of SPE (see Hewlett 1981 for a survey of different approaches to these rules).

247

of the phonetic component is the output of the postlexical phonology.8 According to Port & OTDell's (1985) proposal, FOD would not be dealt with by die phonology, so that FOD obstruents would proceed without change to the phonetic component, where the syllable implementation rule would initiate the correct gesture. Intriguingly, however, this is not the approach adopted by those phonologists who have developed an analysis of FOD in the theoretical framework in which Pulleyblank operates, that of Lexical Phonology. Wiese (1988: 80f.), Rubach (1990) and Hall (1989a, b, 1992b: 53) account for FOD in terms of a phonological rule (a postcyclic rule in Rubach's and Hall's case and a postlexical rule in Wiese's). In other words, the currently available Lexical Phonology approaches to FOD are beset by the very problems which led to Port & ODell's proposals about implementation rules in the first place. Both underlyingly voiced and underlyingly voiceless obstruents enter the phonetic component with a [- voice] specification, which, counterfactually, predicts identical realisations for both in FOD environments. For the present discussion of implementation rules, however, choices made for existing analyses (such as Rubach's, Hall's and Wiese's,) are not crucial. What really matters is whether it is possible in principle to deal with FOD in the phonetic (as opposed to the phonological) component. Consider the following data set taken from Rubach 1990 (p. 84), which was first shown in 2.4.4.2 and is repeated here as (5) for convenience. (5)

Voiced obstruent

Voiceless obstruent

Handl+ung 'act' (handel+n 'to act') Ordn+ung 'order' (ordn+en 'regulate') ebn+en 'flatten' (eben 'flat') Begegn+ung 'meeting' (begegn+en 'meet') eign+en 'own' (eign+en 'to own') nebl+ig 'foggy' (Nebel 'fog')

hand+lich "handy' Bild+nis 'portrait' Ergeb+nis 'result' Wag+nis *boldness' Zeug+nis 'testimony' glaub+lich "believable'

For the correct Hochlautung pronunciations to be arrived at, the obstruent + sonorant clusters in the left-hand column have to be syllabified into a syllable onset. This is problematic for clusters such as [dl], [dn] and [bn] (as pointed out by Vennemann 1968)9

8

9

See also Mohanan 1986 (chapter 6) for a much more detailed discussion of issues associated with phonetic implementation. See 2.4.2.1.2.3 for a detailed discussion of this point.

248

to the extent that these clusters never occur word-initially in this dialect. Several authors, including Wiese 1988 (pp. 89f.), Hall 1992a (p. 71) set up syllable structure conditions which exclude these clusters from the inventory of well-formed syllable onsets, but nevertheless permit such clusters at later stages of derivation. As already mentioned, turning off constraints is standard practice in Lexical Phonology, but it is certainly not the strongest position which could be adopted here. Adherence to the strong position which states that constraints are in force throughout the derivation would produce incorrect results, though. Forms such as Handlung would leave the phonological component with a syllable-final obstruent. The implementation rule proposed by Port & O'Dell takes such syllable-final obstruents as its input and hence initiates the appropriate gesture for devoiced obstruents in these words. This, however, is wrong for Handlung, Ordnung, ebnen and so forth, where FOD does not actually apply in Hochlautung. In other words, implementation rules which take their input from the output of the phonology generate incorrect forms for the pronunciation of a number of German words in the standard variety, unless the phonology permits clusters such as [dn], [dl] or [bn] to be treated as syllable onsets.10 Having discussed two issues arisingfromthe experimental studies of FOD, which are essentially independent of the original question to be answered, let me now return to that original question and see what conclusions were drawn with regard to neutralisation.

5.3.4 Experimental studies and phonological neutralisation Given that most of the experimental studies were based on Kiparsky's 1976 definition of neutralisation, it is clear that those which reported differences between underlyingly voiced and voiceless obstruents in FOD environments could only conclude one thing: FOD was not a neutralisation process. In fact, even by Trubetzkoy's broader classification, FOD would have lost its status as a neutralisation process. Dinnsen (1985) actually went so far as to argue that not only was FOD not a neutralisation process, but that all putative neutralisation processes turned out to be nonneutralising under closer scrutiny. The consequences of this for the sort of phonological theory which depended on neutralisation as a theoretical construct could have been serious. However, Dinnsen's claims seem to have had no perceptible effect on phonological debate. Part of the reason for this may have been the fact that, by the time Dinnsen made his claims, neutralisation by that name was already playing a substantially reduced role in mainstream phonological thinking (see Kaisse & Shaw 1985: 24f.). Also,the notion of Structure Preservation (which had been introduced in Kiparsky 1982b) was

10

The strategy of treating clusters such as [dn], [dl] and [bn] as syllable onsets is adopted, with more or less substantial reservations, by the vast majority of authors (with the possible exception of Lombardi (1991)), including Vennemann (1972), Féry (1991), Giegerich (1992), and Yu (1992b: 37ff).

249

being used to capture what had previously been handled by neutralisation (segmental structure), and possibly even more than that (prosodie structure; see footnote 2). As far as Dinnsen's observations were concerned, though, the new Lexical Phonology broom of Structure Preservation swept only slightly better than the old one of neutralisation. The implication would have been that Structure Preservation was limited to prosodie structure only (if that). This, of course, would have meant substantial erosion (or even complete wiping out) of this theoretical construct. In fact, a considerable amount of evidence has come to light in recent years which suggests that Structure Preservation is much more limited than previously assumed (see, for example, Borowsky 1989, Hall 1989b and Harris 1987, 1989) - something which could be interpreted as a vindication of Dinnsen 1985. So, work in phonology could perhaps have benefitted from taking on board some of the concerns arising from phonetic research.

5.3.5 Summary To summarise, experimental studies which set out to establish whether FOD was a neutralisation process came to the conclusion that it was not, a fact which, when taken together with similarfindingsfor other alleged neutralisation processes, could have had important repercussions for phonological theory, but which was generally ignored in phonological debate. These studies also found evidence that there is a mismatch between production and perception, with speakers being largely unable to reliably perceive distinctions which other members of Iheir speech community (and presumably they themselves) produced. This problem was seen as a strong argument in favour of the dual-lexicon hypothesis, a highly controversial issue which, on the whole, has not been addressed in the phonological literature. Finally, in the light of the fact that the predictions made by the phonology were not borne out by the production facts, some researchers suggested that special implementation rules should handle FOD. This meant that FOD was thus interpreted as a phonetic, as opposed to a phonological, process. However, as I have shown, it does not appear to be possible to accommodate FOD in such a system, at least not when employing the particular syllabifications and rules discussed in 5.3.3. In the next section, I will contrast the analyses available in the SPE-derivative frameworks discussed so far with my GP analysis of FOD, as presented in chapters 3 and 4. It will become apparent that some of the problems just summarised are artifacts of particular theoretical frameworks and that the remainder can be resolved in the framework of GP.

250

5.4 The GP analysis 5.4.1 The proposals As discussed in chapters 3 and 4,1 propose to treat FOD as an effect of autosegmental licensing. Specifically, I have argued that the laryngeal element L (like A) cannot be granted an a-licence by an onset position which has markedly depleted a-licensing potential. There is variation in how different dialects of German deal with the depletion of a-licensing potential. In German Hochlautung and in NSG, for example, onsets which inherit their a-licensing potential from parametrically licensed domain-final empty nuclear positions are unable to license L. Onsets inheriting their a-licensing potential from domain-internal licensed empty nuclei, by contrast, are deemed to have sufficient a-licensing potential for accommodating L in Hochlautung but not in NSG (see 4.4 for detailed discussion). In Hochlautung, then, L is licensed only when preceding a filled nuclear position or an empty nucleus p-licensed through interonset licensing. This means that voiced obstruents (i.e. those which contain a licensed L) only surface next to vowels (e.g. in Rade ["raída], 'wheel' (dat. sg.)) or next to a stem-final consonant which is separated from the FOD obstruent by an empty nucleus (e.g. in ebnen ['?e:bnsn], 'to level') in Hochlautung. In NSG, by contrast, voiced obstruents occur only in the former but not the latter environment. The segmental representations of FOD obstruents, first shown in a slightly different form in 3.4.3.2.1, are repeated here as (6), with an unlicensed element appearing in brackets, following the practice introduced in chapter 4. (6) U

U

@ @ ?I 1 h I 1 L [b]

?I 1 h I 1 (L)

?I 1 h I 1 L [d]

?I 1 h I 1 (L)

I

I

I

I

I

I

L

(L)

? h

u

u

?

h

1 Λ

h I 1 L

h I 1 (L)

M [f] [z] [s] [g] Μ These representations, which contain L (licensed or unlicensed) throughout, contrast with those of non-alternating voiceless obstruents, that is, those which normally appear as p, t, k, fand β in the German orthography. The latter are characterised by the laryngeal element H, which is not sensitive to fluctuations in a-licensing potential and is, therefore, always licensed, regardless of the location along the licensing path of the onset to which [p]

[t]

251

it is attached. The relevant segmental representations are repeated here from 3.4.3.2.1 as (7), for convenience. (7) J

J @ I ?1

?I 1 h I 1 H

I h ι 1 H

I 1 h ι 1 H

i

h ι I H

[Ρ]

Μ

Μ

[f]

[s]

The transcriptions used in (6) and (7) suggest that the voiceless segments are completely indistinguishable from one another phonetically, in spite of the fact that their phonological representations differ. (Those which are underlyingly voiceless contain H, while devoiced obstruents have an unlicensed L.) This is not what is intended. It is simply an artifact of the use of IP A symbols. On the contrary, my analysis predicts that devoiced obstruents (i.e. those with unlicensed L) differ from underlyingly voiceless obstruents (i.e. those with H) at all levels of derivation, including the final output. In other words, my analysis implies that FOD is not a process of neutralisation in the sense of Kiparsky 1976 (and Trubetzkoy 1939 [1969]). If this is true, then both phonological and phonetic evidence should be available to support my claim. Indeed, such evidence can be found.

5.4.2 Phonological evidence 5.4.2.1 Spirantisation of velars: two analyses The phonological evidence comes from NSG, a dialect which, as already mentioned, is very similar to Hochlautung, as far as FOD is concerned. Speakers of both dialects use FOD in identical ways, except for the fact that in Hochlautung the application of FOD is restricted to environments where the segmental material in question is indirectly a-licensed by parametrically licensed domain-final empty nuclear positions, while in NSG any kind of licensed empty nucleus fails to indirectly a-license L. This difference affects only words containing an FOD obstruent followed by a licensed empty nuclear position which is, in turn, followed by a tautomorphemic sonorant consonant. Another difference between the two dialects concerns the spirantisation of devoiced velar stops in NSG. Hochlautung has no truly general spirantisation processes to speak

252

of,11 whereas, under certain circumstances, NSG speakers systematically spirantise all those underlyingly voiced velar stops which occur in FOD environments. Spirantisation, unlike the application of FOD, seems particularly sensitive to the degree of formality accompanying the utterance in question. Some of my informants, for example, spirantise practically all underlyingly voiced velar plosives in FOD environments when producing rapid/casual speech. In a more formal speaking style, such as that used for the recordings on which part of my analysis is based, the same speakers, however, spirantise in a much smaller proportion of possible environments or choose not to spirantise at all (as was the case for one female informant). Investigating this phenomenon would be beyond the brief of this book, so, for the purposes of the present discussion, I shall simply assume that spirantisation occurs consistently for NSG speakers, bearing in mind that this is something of a simplification of the facts. In any case, the main point I want to make here is not affected by the variation just described. The point is that spirantisation can only ever apply to underlyingly voiced velar plosives which occur in FOD environments, whereas underlyingly voiceless velar plosives are never subject to spirantisation. This is illustrated in (8). The contrast between the NSG pronunciation of devoiced velar plosives12 and the corresponding Hochlautung pronunciation is shown in (8a). The absence of spirantisation from underlyingly voiceless velar stops makes the relevant NSG and Hochlautung pronunciations identical, as exemplified in (8b). a. Berg Tag trug Sog lag Sieg Weg Zeug

'mountain' 'day' '(he/she/it) carried' 'suction' '(he/she/it) lay' 'victory' 'path' 'stuff

NSG

Hochlautung

[beçç] [tax] [tru:x] [zo:x] [la:x] [zi:ç] [ve:ç] [tsoiç]

[beçk] [ta:k] [truik] [zo:k] [la:k] [zi:k] [ve:k] [tsoik]

" But see e.g. Wurzel 1981b: 957f. and Hall 1989b, 1992b (chapter 5) for accounts of the spirantisation of the velar stop in the orthographic -ig sequence in Hochlautung, which may either constitute an adjectival suffix or form part of the stem of a small number of nouns. It seems veiy odd, though, that this process should apply only to underlyingly voiced velar plosives which occur after [i]. No other vowel environment (not even [i:]) gives rise to spirantisation. As already mentioned in 4.4.2.1, the issue of whether underlying /g/ is lenited [x] or [ç] is essentially the same as the question of how orthographic is realised. There is a substantial literature dealing with this point, including, more recently, Hall 1989b, 1992b (chapter 5), Yu 1992b (pp. 205218), Macfarland & Pierrehumbert 1991 and Iverson & Salmons 1992.

253

b.

JVSG/Hochlautung

Werk schrak buk Pik

'factory' '(he/she/it) shrank (back)' '(he/she/it) baked' 'spades' (cards)

[veçk] [Jra:k] [bu:k] [pi:k]

Now, if the claim that these spirantisation factsfromNSG support my analysis is to have any force at all, then it is incumbent on me to show that my analysis works better than possible alternative accounts of the same facts. A recent account which deals with precisely the facts under discussion here is offered by Hall (1992b: chapter 5). His analysis is couched in the framework of Lexical Phonology and makes crucial reference to the three rules shown in (9). The individual rules are quoted from pp. 229, 225 and 53 of Hall 1992b respectively. (9)

a. g-Spirantisation (Northern German; optional) - son + voice + high

[+ cont] / [- cons]

b. Dorsal Fricative Assimilation (DFA) morpheme

[- cons]

[- son]

κ 11

ROOT

I [+ cont]

I,··"*

PLACE

DORSAL

]0

254

c. Devoicing [- son] -* [- voice] /

]„

Discussing the ordering of these three rules in the context of the derivation of König (['k0:niç], Tdng'), Hall (1992b: 229) observes that 'Devoicing and Dorsal Fricative Assimilation must both follow g-Spirantisation but are unordered with respect to each other'. It is clear why Devoicing and g-Spirantisation must apply in this counterbleeding order. If Devoicing were to apply first, i.e. if the rules applied in a bleeding order, g-Spirantisation would not have access to that part of the derivational history which is crucial for ensuring that only underlyingly voiced velars spirantise, i.e. the [+ voice] specification which is removed by FOD. Spirantisation would, counterfactually, end up being blocked altogether.13 DFA must apply after g-Spirantisation, that is, in a feeding order, because only fricatives can constitute the input to DFA, which ensures that [x] or [ç] emerge in the appropriate environment. Devoicing and DFA, by contrast do not interact, so that they can apply in any order and still generate the correct output. To illustrate die effects of applying the three rules in the order just established, consider the (simplified) derivation of the noun Sog ([zo:x], 'suction') shown in (10). (10)

Sog

/zo:G/ 1. g-Spirantisation 2. Devoicing 3. DFA

ζο:γ zo:X zo:x [zo:x]

The derivation in (10) clearly generates the correct output, [zo:x]. However, it has two disadvantages. Firstly, it is quite abstract in that it generates at least one segment which never surfaces and secondly, it makes crucial reference to extrinsic rule ordering. Let me begin with the former. The application of g-Spirantisation in (10) generates [γ]. This makes the derivation abstract inasmuch as [γ] never surfaces in NSG.14 The only velars this dialect permits are [g] (obviously only in non-FOD environments), [k] and [x]. In other words, there is no surface motivation for deriving [γ], A derivation which generated only [k] and [x], both 13

14

This is not to say that [k] never occurs for B-speakers. As already mentioned, spirantisation is not always applied, in which case [k] would be the right output. However, the point being made here is that a fricative could not be generated at all, if Devoicing were to be ordered before g-Spirantisation. There are, of course, northern German dialects which exhibit [γ], e.g. in sagen (['zaiyan], 'to say'), such as Berlinisch, for example. NSG, however, is not one of them.

255

of which are part of the NSG inventory, would be more constrained. One could object that Hall makes it clear that underspecification plays an important part in his analysis of German, so that this kind of abstractness is built into the framework itself. After all, underspecified segments (such as /X/ in (10), representing a segment which is specified only as [+ cont, + DORSAL, + high]) cannot be realised as such either. Filling in of redundant feature values is required first. However, the fact that Hall has chosen to nail his colours to the underspecification mast does not invalidate my argument. Λ framework which tolerates only derivations which generate exclusively those segments which are actually attested in a particular language is more constrained than one which generates others as well. The second disadvantage of the approach exemplified in (10) is that it relies on extrinsic rule ordering.15 As already mentioned, in order to enable ^-Spirantisation to distinguish velars with an underlying [+ voice] specification from those without, it is necessary to state that g-Spirantisation precedes Devoicing. GP does not countenance extrinsic rule ordering, that is, in this respect, it is more highly constrained than any framework which does. Moreover, the GP analysis of FOD and spirantisation proposed here can overcome the disadvantages of the Lexical Phonology solution described in the preceding paragraphs. To show how this works, let me attempt to recast the ordered-rule approach illustrated in (10) in terms of the GP framework. In GP, spirantisation is to be interpreted as withdrawal of a-licensing from the occlusion element ?.16 To capture the fact that ^-Spirantisation precedes FOD (as required by die Lexical Phonology account just presented), one could propose that only certain segments are subject to the loss of ?. These segments would have to contain L, which would roughly17 correspond to the [+ voice] specification mentioned in Hall's gSpirantisation rule. Obviously, one would also have to specify the environment in which ? could be lost from a segment containing L. For an NSG speaker, this would be the same environment as that where FOD applies, i.e. preceding any licensed empty nuclear position. That, of course, is the very environment where L is unlicensed. In chapter 3,1 argued that an unlicensed element does not manifest itself, either in the pronunciation or in terms of the complexity of the relevant segment. To the extent that this is the right move, the most consistent - and therefore the best - approach in the present context would be to say that an unlicensed L is never visible within a particular position. So, as far as g-Spirantisation (i.e. withdrawal of a licence from ?) is concerned, 15

16 17

Extrinsic ordering, of course, is only one way of giving a particular rule (g-Spirantisation in this case) access to crucial parts of the derivational history of a segment. Alternatively, one could have proposed a global rule. This device, however, has, on the whole, been rejected in the literature (see, for example, Kiparsky 1976). The original proposal that loss o f ? represents spirantisation is due to Harris (1990: 285). This correspondence is very rough indeed, as the area of overlap between [+ voice] and L is quite small. For example, the feature [+ voice] can be associated with vowels, glides and sonorant consonants in a language such as English, whereas L cannot. This is only possible in tone languages, where L indicates a low tone rather than a voiced sonorant.

256

L is not visible either if it is unlicensed. This means that it is impossible to express the conditions under which ? is lost, if one wants to maintain that the loss of ? precedes final devoicing. In other words, a derivation where ^-Spirantisation precedes FOD is impossible, given the GP analysis presented so far. There is, however, an alternative available. This alternative involves assuming that ? is unlicensed in a velar plosive with no visible (i.e. licensed) laryngeal element, in an FOD environment. This implies the opposite ordering, if one wants to think in terms of ordered rules. The fact that L is unlicensed (i.e. that FOD applies) turns a velar plosive into suitable input to the loss of ?. FOD, if you will, then absolutely feeds spirantisation.18 The derivation of Sog illustrating this approach is shown in (11). (11)

a.

Ο

Ν / \ χ χ

\y o

O N χ

χ

ι

zo:k

(L) b.

Ο

I I

χ

ζ

Ν

ι I

/ \χ V

χ

o

(2)

χ

O N

ι

χ

zo:x

(?)

(L) At no stage does the derivation in (11) generate the unattested voiced velar fricative [γ] or the underspecified /X/, which means that it is less abstract than (10). Like the Lexical Phonology approach, it can also handle cases where spirantisation is suppressed. Recall that suppression of spirantisation occurs in more formal styles, i.e. when NSG speakers make an effort to conform to the prestige dialect of Hochlautung (say, in a formal setting

18

This means that FOD is intrinsically ordered before spirantisation.

257

with speakers from different parts of Germany present or in Fremdwörter, such as Magma·, see 4.4.2.4), and these speakers use Pc] (rather than [g]) on those occasions. This can very easily be captured by simply treating (1 lb) as optional. It may be worth noting at this stage that the constraining principle introduced in 3.3.2 which ensures that only one unit is accessed by a phonological process is respected by this account of FOD and spirantisation. Considering that both processes can be shown to occur independently, it is clear that each makes reference to no more than one unit, and that the fact that two elements are unlicensed in [zo:x] is a consequence of this. To recap briefly, I have argued that the occlusion element ? can remain unlicensed in the representation of a velar plosive in an FOD environment, provided that the affected plosive lacks a LARYNGEAL node. This correctly predicts that /g/ in a non-FOD environment (as in Tage (['taiga], 'days'), for example) and /k/ do not spirantise. The claim that a segment without a licensed LARYNGEAL node may well undergo further reduction receives additional support from the fact that spirantisation of such obstruents is attested in languages other than German. Γ-lenition in English has been analysed along these lines (see Harris & Kaye 1990 and Harris 1990), as has the spirantisation of velars and labials in Greek (Stamatoula Pagoni (p. c.)).

5.4.2.2 Why do only velars spirantise? Before leaving the phonological evidence (i.e. the spirantisation of velars in NSG), I would like to address the question of why it should be velars only which respond to fluctuations in a-licensing potential in this particular way. Asking why velars (rather than alveolare or bilabiale) should spirantise in NSG means invoking the notion of segmental strength. There is a general consensus that segments which are most susceptible to weakening processes (such as spirantisation) tend to be the weakest segments in a particular system or, perhaps, even universally. To express this and other properties which suggest that some segments are stronger than others, phonologists working in the widest conceivable range of theoretical frameworks have set up scales or hierarchies of phonological strength. Such hierarchies can be arrived at in various ways, and some of them take both major class membership and place of articulation of the relevant segments into account.19 As far as major class membership is concerned, there seems to be a considerable amount of common ground. Obstruents are invariably considered as stronger than sonorants (with nasals being stronger than liquids) which, in turn, are treated as having greater phonological strength than vowels.20 Escure's hierarchy of major class and maimer features (see 3.4.3.1) is an example of such a strength scale. Discordant voices, however, make

19

20

Examples of such scales can be found in Vennemann 1972,1982a, Lass 1971, Steriade 1982 (p. 98) and Sigurd 1955. See 2.3 for further references concerning strength scales or sonority hierarchies as well as work containing critical discussion of the fundamental concepts involved.

258

themselves heard when one gets down to finer details and/or makes place of articulation the crucial parameter on the scale. To keep this section as brief as possible, I will not reopen the discussion of problems with such scales which were mentioned earlier in the book (in 3.4.3.1). Instead, I will home in on a particular aspect of strength scales, namely place of articulation, and investigate how well the claims made by hierarchies based on this particular parameter match the facts of human language. Finally, I will discuss the implications which the results of this investigation have for my analysis. Let me begin with Foley (1977), whose strength scales are among the most detailed, most elaborate and well argued I am aware of, although one may disagree with some of the premises they are based on. He sets up a scale (his scale of relative phonological strength a ) which 'refers to the propensity to spirantization, with the weakest element being most inclined to spirantization' (op. cit. : 28). This scale is reproduced in (12) below.21 (12)

g 1

d 2

b 3

It expresses Foley's contention that 'lenition of consonants typically starts with the velar consonant and then generalizes to other consonants' (op. cit. : 27f.). According to Foley, then, there is an implicational hierarchy which says that we can expect to find systems where only velars lenite, others where velars, bilabials and coronals weaken, but none where, say, only coronals undergo this process. The North German dialect we are interested in here fit this pattern perfectly, as do, according to Foley (op. cit.: 3If.), Buriat Mongolian (Altaic; CIS), Czech, Sanskrit, Kasem (Niger-Congo; Ghana, Burkina Faso), Modern Greek, Danish, Spanish and French. Further support for Foley's assessment is provided by Escure's (1977) universal hierarchy of cavity features, which identifies velars as the weakest of the three segment types considered by Foley.22 This view is echoed by Kenstowicz (1981), who quotes Chen's (1975) observation that 'weakening of consonants [...] seems to invariably proceed from the back of the oral cavity towards the front so that velars are most susceptible to weakening [...]' (p. 433). Ni Chasaide (1985: 325) speculates that the comparative sluggishness of the articulator responsible for velars (the tongue body) is likely to make velar segments particularly susceptible to the effects of temporal compression (i.e. a reduction in the amount of time

21

22

The particular α-scale shown here is the first he introduces. He later refines the notion of phonological strength further and identifies this scale as the correct α-scale for Romance languages (pp. 49fl), while the appropriate scale for Germanic languages (based on evidence from strengthening) has coronals and labials appearing in reverse order. The point which is crucial to the present discussion, i.e. that velars are weakest, however, remains unaffected by this refinement. She does not acknowledge any debt to Foley here, but the fact that she uses exactly the same example words from German to illustrate her claim as he does in his 1977 work suggests to me that the two statements are not entirely independent of one another.

259

available for the articulation of a particular segment). Target undershoot, which manifests itself as spirantisation in the case of velar plosives, is the expected result. So far, the consensus seems to be that velars are indeed universally the weakest of all obstruent stops and hence most likely to weaken further. Not everyone, however, agrees with this view. Other workers in the field, many of them critical of the fundamentals of Foley's approach, have adduced a considerable amount of evidence to expose this claim as false. Lass & Anderson (1975: 184) point out that Proto-Uralic *-k- remains unchanged in Hungarian, while *-p- and *-t- are weakened. At the same time, although they are strong intervocalically, velars weaken in initial position, so that Hungarian exhibits two different positional strength hierarchies. Smith (1981) mentions a Danish dialect (S.E. Scania) where g apparently only lenites under certain conditions, while b is invariably weakened to [v], and Harris (1985: 81) cites Viennese as another example of a dialect which has weakened b (to [β]), but retains g. The fact that t is often the only segment to lenite in a number of English dialects (including London and New York city speech; see Harris 1990, Harris & Kaye 1990) could be interpreted as evidence for the claim that coronals, not velars, are weakest there. Finally, according to Hyman (1975: 167), Luganda (Niger-Congo), has intervocalic weakening of all its obstruent stops, with the sole exception of g. I have considered only some of die evidence concerning the status of velars. Yet, even this limited evidence is contradictory. One can infer from this that any attempt to set up a universal strength hierarchy on the basis of place of articulation is probably doomed, a view which is explicitly adopted by, for example, Clements (1990).23 Perhaps one can go as far as to say that certain tendencies exist within language groups, e.g. that velars tend to be weakest and coronals strongest in Germanic languages - but exceptions are always possible. The fact of the matter is probably that 'actual hierarchies by position are language-specific' (Lass & Anderson 1975: 184). In the light of these facts it then seems only appropriate that the framework of GP should be in a position to capture common lenition trajectories (i.e. strength hierarchies) in terms of segmental complexity, without making predictions about place of articulation in this context. Let me now return to the second type of evidence in support of my analysis of FOD as a non-neutralising process, the experimental evidence.

23

It is probably reasonable to assume that authors who set up strength scales, but make no reference to place of articulation implicitly assume that this parameter has no recognisable role to play in this context. The authors who take this view include Lass (1984: 178), Hooper (1976: 206), Selkirk (1984), Wiese 1988 (p. 91) and Hall (1992a, b).

260

5.4.3 Experimental evidence The experimental evidence which supports my analysis of FOD in the framework of GP was presented in 5.3.1, so I will give only a very brief summary of the most important points here. A number of production studies found that there are statistically significant temporal differences between devoiced obstruents and their underlyingly voiceless congeners. Perception studies carried out in conjunction with the production studies also showed that hearers are able to distinguish between these two segment types, but only to a limited extent. The proportion of correctly identified tokens was consistently better than chance, but it is unclear whether the distinction is sufficiently salient for it to play a crucial role in normal communication.

5.4.4 The general implications for earlier problems Some of the problems with the orthodox analysis of FOD as [- son] [- voice] appear to have been resolved by the new analysis in the GP framework. The conflict between the phonological predictions and the production facts no longer exists. The phonological analysis itself makes the claim that underlyingly voiced obstruents differ from their underlyingly voiceless congeners even in FOD environments. In other words, there is no need for this to be handled by phonetic implementation rules. The general consensus among phonologists is that FOD is a phonological event, and it can be treated as such in GP, without predicting that absolute neutralisation results from it. This, however, does not mean that the mismatch between production and perception has disappeared, too. The silver lining of this cloud is that, unlike practically all phonological theories available today, GP does not rely on primarily articulalory categories to define the phonological elements it employs (see 3.3.1 for a brief discussion). Fully fledged acoustic definitions are available for all elements, and the articulatory labels used in this book are only a matter of the reader's convenience. The GP approach maintains the original Chomskyan (1965: 3) assumption that linguistic theory and hence the grammar developed by linguists is neutral with regard to the speaker-hearer. At the same time, though, it captures the fact that perception precedes production in language acquisition and that it is with reference to cognitive categories based on sound patterns that the child makes sense of what she hears. By implication, the prediction is made that the difference between phonologically distinct units has to be perceptually salient for a child to be able to acquire it. The only conceivable exceptions to this must either themselves be part of UG or be deducible from evidence provided by alternations. For example, it may not be necessary for the distinction between a governed obstruent and a governing obstruent to be perceptually salient, as complexity requirements are contained in UG. Also, some phonological differences may not be perceptually salient in certain environments or may only be salient to a limited extent (such as the difference between a devoiced obstruent and its

261

underlyingly voiceless congener in an FOD environment), but the child will be able to establish the phonological distinction by reference to alternation facts together with her innate phonological 'tool-kit', which, of course, is also part of UG. The specific claim for FOD which I want to make here is that FOD belongs to UG. It may be the case that reduced a-licensing potential typically results in weakening, where the LARYNGEAL node is one of the prime candidates for losing its licence. One could speculate that there is a parameter which controls how sensitive languages are to fluctuations in a-licensing potential. Some, which appear to lack segmental reduction processes (French, for example) seem completely insensitive to it, whereas others, such as German, respond to such fluctuations with segmental reduction events. If it is indeed a parameter which is involved here, it would be reasonable to assume that its default setting is such that reduced a-licensing potential is both registered and acted on. The depletion of a-licensing potential is particularly marked (in the non-technical sense) where the Final Empty Nucleus Parameter (see 4.2.1) is set at ON and provides an additional 'dose' of p-licensing to certain domain-final empty nuclear positions. This claim is supported by the facts referred to in 1.1, which included the following observations. Children, when learning a language with domain-final empty nuclear positions, go through a stage of applying FOD (usually at the point when final obstruents are first acquired), even in languages where adult speakers do not. Similarly, native speakers of languages which do not parametrically license final empty nuclei characteristically apply FOD to foreign words with final obstruents (i.e. words from a language with final empty nuclei), even if the source language does not exhibit FOD (see Stampe 1969 (p. 445) for both these points).

5.5 Conclusion In this chapter I have briefly highlighted parts of the history of the theoretical construct of neutralisation and examined the role it has played in experimental work over the past decade or so. I have discussed some issues arising from this experimental work which pose problems for those phonological theories which sparked off heightened interest in the phonetic aspects of neutralisation. By relating these problems to my GP analysis I have been able to show that it is possible to overcome them. In fact, it seems that neutralisation as a theoretical construct is rapidly disappearing from the phonological landscape. In Lexical Phonology, it has been replaced by Structure Preservation (which, itself is being exposed as much less far-reaching and important than originally assumed) and in GP it has no theoretical status at all. That is not to say, though, that the word 'neutralisation' will soon be missing from phonologists' vocabulary. It is still a useful cover term for a wide range of effects, whatever their theoretical status may be.

References

The following abbreviations have been used throughout: CIS JASA JL JPh LB Lg LI NELS NLLT ZS

Papers from the Annual Regional Meeting, Chicago Linguistic Society Journal of the Acoustical Society of America Journal of Linguistics Journal of Phonetics Linguistische Berichte Language Linguistic Inquiry Proceedings of the Annual Meeting of the North Eastern Linguistic Society Natural Language and Linguistic Theory Zeitschrift für Sprachwissenschaft

Ali, Latif, Ray Daniloff & Robert Hammarberg (1979). Intrusive stops in nasal-fricative clusters: an aerodynamic and acoustic investigation. Phonetica 36. 85-97. Anderson, John M. & Colin J. Ewen (1987). Principles of Dependency Phonology. Cambridge: Cambridge University Press. — & Jacques Durand (1986). Dependency Phonology. In Durand (1986b). 1-54. Archangeli, Diana (1984). Underspecification in Yawelmani phonology and morphology. PhD dissertation. Cambridge, MA: MIT. Published 1988, New York: Garland. — (1988). Aspects of Underspecification Theory. Phonology 5. 183-207. Aronoff, Mark & Richard T. Oehrle (eds.) (1984). Language sound structure: studies in phonology presented to Morris Halle by his teacher and students. Cambridge, ΜΑ: ΜΓΓ Press. Avery, Peter & Keren D. Rice (1989). Segment structure and coronal underspecification. Phonology 6.179-200. Bach, Adolf (1961). Geschichte der deutschen Sprache. 7th edn. Heidelberg: Quelle & Meyer. Backley, Phillip (1993). Coronal: the undesirable element. VCL Working Papers in Linguistics S. 301-323. Barry, Susan M.E. (1989). Aspects of sonorant devoicing in Russian. Speech, Hearing and Language: work in progress 3. London: University College London. 47-59. Benware, Wilbur A. (1986). Phonetics and phonology of modern German: an introduction. Washington, DC: Georgetown University Press. Berendsen, Egon (1983). Final devoicing, assimilation, and subject clitics in Dutch. In Hans Bennis & W.U.S. van Lessen Kloeke (eds.) Linguistics in the Netherlands 1983. Dordrecht: Foris. 21-29. — (1987). Schwa deletion in Dutch cliticization. In Dressier et al. (1987). 13-20. Booij, Geert E. & Jerzy Rubach (1987). Postcyclic versus postlexical rules in Lexical Phonology. LI 18. 1-44. Borowsky, Toni (1989). Structure preservation and the syllable coda in English. NLLT 7. 145-166. Broadbent, Judith M. (1991). Linking and intrusive r in English. VCL Working Papers in Linguistics 3. 281-301. Brockhaus, Wiebke G. (1990). Colourful leagues: a Government Phonology approach to final obstruent devoicing in German. VCL Working Papers in Linguistics 2. 270-297. — (1991). Final devoicing and neutralisation. VCL Working Papers in Linguistics 3. 303-330.

264 — (1992a). Final devoicing: empty nuclei and element-licensing. Paper presented at the Government Phonology Workshop, 7th International Phonology Meeting, Krems (Austria). — (1992b). Final devoicing: principles and parameters. PhD dissertation. London: UCL, University of London. — (1993). Autosegmental licensing and Licensing Inheritance in German. Paper presented at the Cognitive Phonology Workshop, Manchester. — (1994). Segmental representations without [R]? Some evidence from German. Paper presented at the Government Phonology Workshop, Vienna. — (1995). Skeletal and suprasegmental structure within Government Phonology. In Durand & Katamba (1995). 180-221. Bruck, Anthony, Robert A. Fox & Michael W. La Galy (eds.) (1974). Papers from the parasession on Natural Phonology. Chicago: Chicago Linguistic Society. Butt, Matthias (1992). Sonority and the explanation of syllable structure. LB 137. 45-67. Carr, Philip (1993). Phonology. Basingstoke: Macmillan. Chambers, W. Walker & John R. Wilkie (1970). A short history of the German language. London: Methuen. Charette, Monik (1988). Some constraints on governing relations in phonology. PhD dissertation. Montreal: McGill University. — (1989). The Minimality Condition in phonology. JL 25.159-187. — (1990). Licence to govern. Phonology 7. 233-253. — (1991). Conditions on phonological government. Cambridge: Cambridge University Press. — (1992). Mongolian and Polish meet government-licensing. SOAS Working Papers in Linguistics and Phonetics!. 275-291. — & AsL Göksel (1994). Switching and vowel harmony in Turkic languages. SOAS Working Papers in Linguistics and Phonetics 4. 31-52. Charles-Luce, Jan (1985). Word-final devoicing in German: effects of phonetic and sentential contexts. JPh 13. 309-324. — & Daniel A. Dinnsen (1987). A reanalysis of Catalan devoicing. JPh 15. 187-190. Chen, Matthew (1970). Vowel length variation as a function of the voicing of the consonant environment. Phonetica 22. 129-159. — (1975). Metarules and universal constraints in phonological theory. Proceedings of the 11th International Congress of Linguists. Bologna, 1972. 1152-1167. Chomsky, Noam (1965). Aspects of the theory of syntax. Cambridge, MA: MIT Press. — & Morris Halle (1968). The Sound Pattern of English. New York: Harper & Row. Clements, George N. (1985). The geometry of phonological features. Phonology Yearbook 2. 225-252. — (1987). Phonological feature representation and the description of intrusive stops. CLS 23, Part 2 : Parasession on Autosegmental and Metrical Phonology. 29-50. — (1990). The role of the sonority cycle in core syllabification. In John Kingston & Mary E. Beckman (eds.) Papers in Laboratory Phonology I: between the grammar and physics of speech. Cambridge: Cambridge University Press. 283-333. — & Elizabeth V. Hume (1993). The internal organization of speech sounds. Ms. Paris and Columbus, Ohio: I.L.P.G.A. and Ohio State University. — & Samuel Jay Keyser (1983). CV phonology: a generative theory of the syllable. Cambridge, MA: ΜΓΓ Press. Coleman, John (1992). The phonetic interpretation of headed phonological structures containing overlapping constituents. Phonology 9. 1-44. — (1995). Declarative Lexical Phonology. In Durand & Katamba (1995). 333-382. Crystal, David (1987). The Cambridge encyclopedia of language. Cambridge: Cambridge University Press.

265 Cyran, Eugeniusz (1995). Vocalic elements in phonology: a study in Munster Irish. PhD dissertation. Lublin: Catholic University of Lublin. Davidsen-Nielsen, Niels (1978). Neutralization and archiphoneme: two phonological concepts and their history. Copenhagen: Akademisk Forlag. Davis, Stuart (1988). Syllable onsets as a factor in stress rules. Phonology 5.1-19. Dialo, A. (1981). Une phonologie du wolof. Ms. Dakar: Centre de linguistique appliquée de Dakar. — (1983). Eléments systématique du wolof contemporain. Ms. Dakar: Centre de linguistique appliquée de Dakar. Dikken, Marcel den & Harry G. van der Hulst (1988). Segmental hierarchitecture. In van der Hulst & Smith (1988: vol. 1). 1-78. Dinnsen, Daniel A. (1980). Phonological rules and phonetic explanation. JL16.171-191. — (1985). A re-examination of phonological neutralization. JL 21. 265-279. — & Jan Charles-Luce (1984). Phonological neutralization, phonetic implementation and individual differences. JPh 12. 49-60. Dressier, Wolfgang U., Hans C. Luschützky, Oskar E. Pfeiffer & John R. Rennison (eds.) (1987). Phonologica 1984: Proceedings of the Fifth International Phonology Meeting, Eisenstadt, 25-28 June 1984. Cambridge: Cambridge University Press. Drosdowski, Günther, Rudolf Köster & Wolfgang Müller (eds.) (1976-1981). DUDEN: Das große Wörterbuch der deutschen Sprache. 6 vols. Mannheim: Bibliographisches Institut —.Wolfgang Müller, Werner Scholze-Stubenrecht & Matthias Wermke (eds.) (1991). Duden 1: Rechtschreibung der deutschen Sprache. 20th edn. Mannheim: Bibliographisches Institut. Durand, Jacques (1986a). French liaison, floating segments and other matters in a Dependency framework. In Durand (1986b). 161-201. — (ed.) (1986b). Dependency and non-linear phonology. London: Croom Helm. — (1990). Generative and non-linear phonology. Harlow: Longman. — & Francis Katamba (eds.) (1995). Frontiers of phonology: atoms, structures, derivations. Harlow: Longman. Eisenberg, Peter, Karl Heinz Ramers & Heinz Vater (eds.) (1992). Silbenphonologie des Deutschen. Tübingen: Gunter Narr. Escure, Geneviève J. (1975). Weakening and deletion processes in language change. Unpublished PhD dissertation. Bloomington: Indiana University. — (1977). Hierarchies and phonological weakening. Lingua 43. 55-64. Fay, David & Anne Cutler (1977). Malapropisms and the structure of the mental lexicon. LI 8. 505-520. Féry, Caroline (1991). German schwa in prosodie morphology. ZS 10. 65-85. Fleischer, Wolfgang (1975). Wortbildung der deutschen Gegenwartssprache. 4th edn. Tübingen: Niemeyer. — & Irmhild Barz (1992). Wortbildung der deutschen Gegenwartssprache. Tübingen: Niemeyer. Foley, James (1977). Foundations of theoretical phonology. Cambridge: Cambridge University Press. Fourakis, Marios (1984). Should neutralization be redefined? JPh 12. 291-296. — & Gregory K. Iverson (1984). On the 'incomplete neutralization' of German final obstruents. Phonetica 41. 140-149. — & Robert F. Port (1986). Stop epenthesis in English. JPh 14.197-221. Fox, Anthony (1990). The structure of German. Oxford: Clarendon Press. Giegerich, Heinz J. (1985). Metrical phonology and phonological structure: English and German. Cambridge: Cambridge University Press. — (1987). Zur Schwa-Epenthese im Standarddeutschen. LB 112. 449-469. — (1989). Syllable structure and lexical derivation in German. Indiana University Linguistics Club. — (1992). Onset maximisation in German: the case against resyllabification rules. In Eisenberg et al. (1992). 134-171. Goldsmith, John A. (1989). Licensing, inalterability and harmonic rule application. CLS 25.145-157.

266 — (1990). Autosegmental and Metrical Phonology. Oxford: Blackwell. Goyvaerts, Didier L. (1981a). Introduction. In Goyvaerts (1981b). 1-26. — (ed.) (1981b). Phonology in the 1980s. Ghent: Story-Scientia. Griffen, Toby D. (1982). German /r/. Lingua 56. 297-316. Grunwell, Pamela (1987). Clinical phonology. 2nd edn. London: Croom Helm. Guerssel, Mohand (1990). On the syllabification pattern of Berber. Ms. Montreal: Université du Québec à Montréal. — & Jean Lowenstamm (1988). Classical Arabie 'metathesis'. Paper presented at the Generative Phonology Workshop, Leiden. Gussmann, Edmund (1992). Resyllabification and delinking: the case of Polish voicing. LI 23. 29-56. — & Jonathan D. Kaye (1993). Polish notes from a Dubrovnik café I: the yers. SOAS Working Papers in Linguistics and Phonetics 3. 427-462. Haggard, Mark, Stephen Ambler & Mo Callow (1970). Pitch as a voicing cue. JASA 47. 613-617. Hall, Tracy Alan (1989a). German syllabification, the velar nasal, and the representation of schwa. Linguistics 27. 807-842. — (1989b). Lexical phonology and the distribution of German [ç] and [x]. Phonology 6.1-17. — (1992a). Syllable final clusters and schwa epenthesis in German. In Eisenberg et al. (1992). 208-245. — (1992b). Syllable structure and syllable-related processes in German. Tübingen: Niemeyer. — (1993). The phonology of German /rJ. Phonology 10. 83-105. Halle, Morris & George N. Clements (1983). Problem book in phonology: a workbook for introductory courses in linguistics and in modern phonology. Cambridge, MA: MIT Press. — & Kenneth N. Stevens (1971). A note on laryngeal features. Quarterly Progress Report of the Research Laboratory of Electronics, Massachusetts Institute of Technology. 101. 198-213. Hammond, Robin (1981). A German reference grammar. Oxford: Oxford University Press. Harris, John (1985). Phonological variation and change: studies in Hiberno-English. Cambridge: Cambridge University Press. — (1986). Phonetic constraints on sociolinguistic variation. Sheffield Working Papers in Language and Linguistics 3. 120-142. — (1987). Non-structure-preserving rules in Lexical Phonology: southeastern Bantu harmony. Lingua 73. 255-292. — (1989). Towards a lexical analysis of sound change in progress. JL 25. 35-56. — (1990). Segmental complexity and phonological government. Phonology 7. 255-300. — (1992). Licensing Inheritance. UCL Working Papers in Linguistics 4. 359-406. — (1994a). English sound structure. Oxford: Blackwell. — (1994b). Monovalency and opacity: Chichewa height harmony. UCL Working Papers in Linguistics 6. 509-547. — (1994c). Nonspecified coronality as weak licensing. Paper presented at the Government Phonology Workshop, Vienna. — & Geoff A. Lindsey (1995). The elements of phonological representation. In Durand & Katamba (1995). 33-79. — & Jonathan D. Kaye (1990). A tale of two cities: London glottalling and New York City tapping. Ms. London: University College London and SOAS. Also in The Linguistic Review 7.3. 251-274. Hayes, Bruce (1982). Extrametricality and English stress. LI 13. 227-276. — (1984). The phonetics and phonology of Russian voicing assimilation. In Aronoff & Oehrle (1984). 318-328. — (1986). Inalterability in CV Phonology. Lg 62. 321-351. Hayward, Katrina (1993). /p/ vs. lb/ in Javanese: some preliminary data. SOAS Working Papers in Linguistics and Phonetics 3. 1-33. Heidolph, Karl Erich, Walter Flämig & Wolfgang Mötsch (eds.) (1981). Grundzüge einer deutschen Grammatik. Berlin: Akademie-Verlag.

267 Heike, Georg (1992). Zur Phonetik der Silbe. In Eisenberg et al. (1992). 1-44. Heo, Yong (1990). Empty nuclei in Middle Korean. SOAS Working Papers in Linguistics and Phonetics 1. 21-37. Hewlett, Nigel (1981). Phonetic realisation rules in generative phonology. JPh 9. 63-77. Hildebrandt, Bruno F.O. & Lieselotte M. Hildebrandt (1965). Das Deutsche R: Regelhaftigkeiten in der gegenwärtigen Reduktions-Entwicklung und Anwendung im Fremdsprachenuntenicht. Linguistics 11. 5-20. Hirose, H., C.Y. Lee & T. Ushijima (1974). Laryngeal control in Korean stop production. JPh 2.145152. Höhle, Tilman Ν. & Heinz Vater (1978). Derivational constraints und die silbischen Konsonanten im Deutschen. In Henrik Birnbaum, Lubomir Ôuroviè & Gunnar Jacobsson (eds.) (1978). Studia linguistica Alexandro VasiliifilioIssatschenko a collegis amicisque oblata. Lisse: Peter de Ridder. 169-186. Hogg, Richard M. & C.B. McCully (1987). Metrical phonology: a coursebook. Cambridge: Cambridge University Press. Hombert, Jean-Marie, John J. Ohala & William G. Ewan (1979). Phonetic explanations for the development of tones. Lg SS. 37-58. Hooper, Joan B. (1976). An introduction to natural generative phonology. New York: Academic Press. Houlihan, Kathleen & Gregory K. Iverson. (1979). Functionally-constrained phonology. In Daniel A. Dinnsen (ed.) Current approaches to phonological theory. Bloomington: Indiana University Press. 51-73. Hulst, Harry G. van der (1984). Syllable structure and stress in Dutch. Dordrecht: Foris. — (1988). The geometry of vocalic features. In van der Hulst & Smith (1988: vol. 2). 77-125. — (1992). On the nature of phonological primes and the structure of segments. Ms. Leiden: Department of General Linguistics, University of Leiden. — & Norval S.H. Smith (1982a). An overview of autosegmental and metrical phonology. In van der Hulst & Smith (1982b: vol. 1). 1-46. — & — (eds.) (1982b). The structure of phonological representations. 2 vols. Dordrecht: Foris. — & — (eds.) (1988). Features, segmental structure and harmony processes. 2 vols. Dordrecht: Foris. Hyman, Larry M. (1975). Phonology: theory and analysis. New York: Holt, Rinehart and Winston. — (1985). A theory of phonological weight. Dordrecht: Foris. Ingleby, Michael, Wiebke G. Brockhaus & Carl R. Chalfont (1994). Integration of linguistic knowledge into pattern recognition. Proceedings of the A1SB Workshop on Computational Linguistics for Speech and Handwriting Recognition. Leeds, April 1994. —, — & — (1994). Robust techniques for recognition of new knowledge-based speech primitives. Proc. Fifth Australian International Conference on Speech Science and Technology. Perth (Australia), December 1994. Issatschenko, Alexander (1974). Das "schwa mobile" und "schwa constans" im Deutschen. In Ulrich Engel & Paul Grebe (eds.) Sprachsystem und Sprachgebrauch: Festschrift für Hugo Moser zum 65. Geburtstag. Düsseldorf: Pädagogischer Verlag Schwann. 142-171. Itô, Junko (1986). Syllable theory in prosodie phonology. PhD dissertation. Amherst, MA: University of Massachusetts. — & R. Armin Mester (1986). The phonology of voicing in Japanese: theoretical consequences for morphological accessibility. U 1 7 . 49-73. Iverson, Gregory K. & Joe Salmons (1992). The place of Structure Preservation in German diminutive formation. Phonology 9. 137-143. Jakobson, Roman (1929). Remarques sur l'évolution phonologique du russe. (= Travaux du cercle linguistique de Prague 2). — (1978). Mutual assimilation of Russian voiced and voiceless consonants. Studia Lingüistica 32.107110.

268 Jassem, Wiktor & Lutoslawa Richter (1989). Neutralization of voicing in Polish obstruents. JPh 17. 317-325. Kagaya, Ryohei (1974). A fiberscope and acoustic study of the Korean stops, affricates and fricatives. JPh 2.161-180. Kaisse, Ellen M. (1986). Locating Turkish devoicing. Proceedings of the West Coast Conference on Formal Linguistics 5.119-128. — & Patricia A. Shaw (1985). On the theory of Lexical Phonology. Phonology Yearbook 2.1-30. Katamba, Francis (1989). An introduction to phonology. Harlow: Longman. Katzner, Kenneth (1986). The languages of the world. 2nd edn. London: Routledge & Kegan Paul. Kaye, Jonathan D. (1987). Government in phonology: the case of Moroccan Arabic. The Linguistic Review 6. 131-159. — (1988). On the interaction of theories of lexical phonology and theories of phonological phenomena. Ms. London: SOAS. Also in Wolfgang U. Dressier, Hans C. Luschützky, Oskar E. Pfeiffer & John R. Rennison (eds.) (1992). Phonologica 1988: Proceedings of the Sixth International Phonology Meeting, Krems, 1-4 July 1988. Cambridge: Cambridge University Press. 141-155. — (1989). Phonology: a cognitive view. Hillsdale, NJ: Lawrence Erlbaum Associates. — (1990a). The strange vowel sets of Charm theory: the question from top to bottom. JL 26.175-181. — (1990b). 'Coda' licensing. Phonology 7. 301-330. — (1992). Do you believe in magic? The story of s+C sequences. SOAS Working Papers in Linguistics and Phonetics 2. 293-313. — (1993). Derivations and interfaces. SOAS Working Papers in Linguistics and Phonetics 3. 90-126. —, Jean Lowenstamm & Jean-Roger Vergnaud (1985). The internal structure of phonological elements: a theory of charm and government. Phonology Yearbook 2. 305-328. —, — &— (1990). Constituent structure and government in phonology. Phonology 7. 193-231. Keating, Patricia A. (1984). Phonetic and phonological representation of stop consonant voicing. Lg 60. 286-319. Kenstowicz, Michael J. (1981). Functional explanation in generative phonology. In Goyvaerts (1981b). 431-444. — (1994). Phonology in generative grammar. Oxford: Blackwell. — & Charles W. Kisseberth (1979). Generative phonology: description and theory. New Yoik: Academic Press. King, Robert D. (1969). Historical linguistics and generative grammar. Englewood Cliffs, NJ: Prentice-Hall Kiparsky, Paul (1973). Elsewhere in phonology. In Stephen R. Anderson & Paul Kiparsky (eds). A Festschrift for Morris Halle. New York: Holt, Rinehart and Winston. 93-106. — (1976). Abstractness, opacity, and global rules. In Andreas Koutsoudas (ed.) The application and ordering of grammatical rules. The Hague: Mouton. 160-186. — (1979). Metrical structure assignment is cyclic. LI 10. 421-441. — (1982a). From cyclic phonology to lexical phonology. In van der Hulst & Smith (1982b: vol. 1). 131176. — (1982b). Lexical Morphology and Phonology. In I.-S. Yang (ed.) Linguistics in the morning calm. Seoul: Hanshin. 3-91. — (1985). Some consequences of Lexical Phonology. Phonology Yearbook 2.85-138. Klatt, Dennis H. (1981). Lexical representations for speech production and perception. In Terry Myers, John Laver & John Anderson (eds.) The cognitive representation of speech. Amsterdam: NorthHolland. 11-31. Kloeke, W.U.S. van Lessen (1982a). Deutsche Phonologie und Morphologie: Merkmale und Markiertheit. Tübingen: Niemeyer. — (1982b). Exteme Argumente in der Sprachbeschreibung. In Vennemann (1982b). 171-182. Kohler, Klaus J. (1977). Einführung in die Phonetik des Deutschen. Berlin: Erich Schmidt Verlag.

269 — (1982). F 0 in the production of lenis and fortis plosives. Phonetica 39.199-218. — (1984). Phonetic explanation in phonology: the feature fortis/lenis. Phonetica 41. 150-174. König, Werner (1978). dtv-Atlas zw deutschen Sprache. Munich: Deutscher Taschenbuch Verlag. Krech, Eva-Maria (1968). Sprechwissenschaftlich-phonetische Untersuchungen zum Gebrauch des Glottisschlageinsatzes in der allgemeinen deutschen Hochlautung. Basel: Karger. Lass, Roger (1971). Boundaries as obstruents: Old English voicing assimilation and universal strength hierarchies. JL 7. 15-30. — (1980). On explaining language change. Cambridge: Cambridge University Press. — (1984). Phonology: an introduction to basic concepts. Cambridge: Cambridge University Press. — & John M. Anderson (1975). Old English phonology. Cambridge: Cambridge University Press. Lea, W.A. (1972). Intonational cues to the constituent structure and phonemics of spoken English. PhD dissertation. Purdue University. Lester, Leland & Royal Skousen (1974). The phonology of drunkenness. In Bruck et al. (1974). 233239. Levin, Juliette (1985). A metrical theory of syllabicity. PhD dissertation. Cambridge, MA: MIT. Lightner, τ.Μ. (1963). A note on the formulation of phonological rules. Quarterly Progress Report of the Research Laboratory of Electronics, Massachusetts Institute of Technology. 68. 187-189. Lindsey, Geoff A. & John Harris (1990). Phonetic interpretation in generative grammar. UCL Working Papers in Linguistics 2. 355-369. Linell, Per (1979). Psychological reality in phonology: a theoretical study. Cambridge: Cambridge University Press. Lisker, Leigh & Arthur S. Abramson (1964). A cross-language study of voicing in initial stops: acoustical measurements. Word 20. 386-422. Lodge, Ken R. (1987). French phonology again: deletion phenomena, syllable structure and other matters. In John M. Anderson & Jacques Durand (eds.) Explorations in Dependency Phonology. Dordrecht: Foris. 133-168. — (1989). A non-segmental account of German Umlaut: diachronic and synchronic perspectives. LB 124. 470-491. Löfqvist, Anders (1975). Intrinsic and extrinsic F 0 variations in Swedish tonal accents. Phonetica 31. 228-247. Löhken, Sylvia C. (1995). The emergence of complex consonantal clusters in German: an OptimalityTheoretical approach. Paper presented at HILP 2, Amsterdam. Lombardi, Linda (1991). Laryngeal features and laryngeal neutralization. PhD dissertation. Amherst, MA: University of Massachusetts. Lowenstamm, Jean (1988). On the non-adjacency of geminate consonants. Paper presented at the Generative Phonology Workshop, Leiden. Macbeth, N. (1974). Darwin retried. London: Gamstone Press. Macfarland, Talke & Janet Pierrehumbert (1991). On ich-Laut, ach-Laut and Structure Preservation. Phonology 8.171-180. Maddieson, Ian (1984). Patterns of sounds. Cambridge: Cambridge University Press. Magee, Bryan (1982). Popper. 2nd edn. London: Fontana. Mangold, Max, Günther Drosdowski, Wolfgang Müller, Werner Scholze-Stubenrecht & Matthias Wermke (eds.) (1990). DUDEN 6: Aussprachewörterbuch: Wörterbuch der deutschen Standardaussprache. 3rd edn. Mannheim: Dudenverlag. Mascaró, Joan (1986). Syllable-final processes in Catalan. In Neidle & Nufiez Cedeño (1986). 163-180. — (1987a). A reduction and spreading theory of voicing and other sound effects. Ms. Barcelona: Universität Autonoma de Barcelona. — (1987b). Underlying voicing recoverability of finally devoiced obstruents in Catalan. JPh 15.183186.

270 Massaro, Dominic W. & Michael M. Cohen (1977). Voice onset time and fundamental frequency as cues to the /zi/-/si/ distinction. Perception & Psychophysics 22. 373-382. McCarthy, John J. (1986). OCP effects: gemination and antigemination. LI 17. 207-263. — (1988). Feature geometry and dependency: a review. Phonetica 45. 84-108. Meinhold, Gottfried & Eberhard Stock (1982). Phonologie der deutschen Gegenwartssprache. 2nd edn. Leipzig: VEB Bibliographisches Institut. Meyer-Eppler, W. (1959). Zur Spektralstruktur der /r/-Allophone des Deutschen. Acustica 9. 247-250. Mohanan, K.P. (1986). The theory of Lexical Phonology. Dordrecht: Reidel. — (1991). On the bases of radical underspecification. NLLT9. 285-325. Moulton, William G. (1962). The sounds of English and German. Chicago: Chicago University Press. Murray, Robert W. & Theo Vennemann (1983). Sound change and syllable structure in Germanic phonology. Lg 59. 514-528. Muthmann, Gustav (1991). Rückläufiges deutsches Wörterbuch: Handbuch der Wortausgänge im Deutschen, mit Beachtung der Wort- und Lautstruktur. 2nd edn. Tübingen: Niemeyer. Myers, Scott (1987). Vowel shortening in English. NLLT S. 485-518. Neidle, C. & R.A. Nuñez Cedeño (eds.) (1986). Studies in Romance languages. Dordrecht: Foris. Nespor, Marina I. & Irene Vogel (1986). Prosodie phonology. Dordrecht: Foris. Ni Chasaide, Ailbhe (1985). Preaspiration in phonological stop contrasts. Unpublished PhD dissertation. Bangor: University College of North Wales. — (1989). Sonorization and spirantization: a single phonetic process? Ms. Dublin: Trinity College. Nolan, Francis (1986). The implications of partial assimilation and incomplete neutralisation. Cambridge Papers in PAonetics and Experimental Linguistics 5. N l - N l l . O'Dell, Michael L. & Robert F. Port (1983). Discrimination of word-final voicing in German. JASA 73, Suppl. 1. S31. O'Kane, Donai (1978). Manner of vowel termination as a perceptual cue to the voicing status of postvocalic stop consonants. JPh 6. 311-318. Ohala, John J. (1974). Phonetic explanation in phonology. In Bruck et al. (1974). 251-274. Pagoni, Stamatia (1993). Modern Greek phonological variation: a Government Phonology approach. PhD dissertation. London: UCL, University of London. Paradis, Carole & Jean-François Prunet (eds.) (1991). The special status of coronals: internal and external evidence. San Diego: Academic Press. Parker, Frank (1974). The coarticulation of vowels and stop consonants. JPh 2. 211-221. — (1981). A functional-perceptual account of final devoicing. JPh 9. 129-137. Penzl, Herbert (1975). Vom Urgermanischen zum Neuhochdeutschen. Berlin: Erich Schmidt. Piggott, Glyne L. (1988). A parametric approach to nasal harmony. In van der Hulst & Smith (1988: vol. 1). 131-167. Port, Robert F. & Penny Crawford (1989). Incomplete neutralization and pragmatics in German. JPh 17. 257-282. — & Michael L. O'Dell (1985). Neutralization of syllable-final voicing in German. JPh 13. 455-471. —, Fares Mitleb & Michael L. O'Dell (1981). Neutralization of voicing in German is incomplete. JASA 70, Suppl. 1. S13. Priebsch, W. & W.E. Collinson (1958). The German language. 4th edn. London: Faber and Faber. Prince, Alan S. (1984). Phonology with tiers. In Aronoff & Oehrle (1984). 234-244. — & Paul Smolensky (1993). Optimally Theory: constraint interaction in generative grammar. RuCCs Technical Report #2. Rutgers University Center for Cognitive Science, Rutgers University. Prunet, Jean-François (1986). Liaison and Nasalization in French. In Neidle & Nuñez Cedeño (1986). 225-236. Pulleyblank, Douglas G. (1986). Tone in Lexical Phonology. Dordrecht: Reidel. Pullum, Geoffrey K. (1976). The Duke of York gambit. JL 12.

271 Pye, Susan M.E. (1986). Word-final devoicing of obstruents in Russian. Cambridge Papers in Phonetics and Experimental Linguistics 5. P1-P10. Ramers, Karl Heinz (1992). Ambisilbische Konsonanten im Deutschen. In Eisenberg et al. (1992). 246283. Rennison, John R. (1981). What is shwa in Austrian German? The case for epenthesis, and its consequences. In Wolfgang U. Dressier, Oskar E. Pfeiffer & John R. Rennison (eds.) Phonologica 1980: Akten der Vierten Internationalen Phonologie-Tagung, Wien, 29. Juni - 2. Juli 1980. Innsbruck: Innsbrucker Beiträge zur Sprachwissenschaft. 351-356. — (1993). Empty nuclei in Koromfe: a first look. Ms. Vienna: University of Vienna. Rice, Keren D. (1992). On deriving sonority: a structural account of sonority relationships. Phonology 9. 61-99. — & Peter Avery (1990). On the representation of voice. NELS 20. 428-442. Rischel, J0rgen (1991). The relevance of phonetics for phonology: a commentary. Phonetica 48. 233-262. Roca, Iggy (1994). Generative phonology. London: Routledge. Rubach, Jeizy (1982). Analysis of phonological structures. Warsaw: Panstwowe Wydawnictwo Naukowe. — (1984). Cyclic and Lexical Phonology: the structure of Polish. Dordrecht: Foris Publications. — (1990). Final devoicing and cyclic syllabification in German. LI 21. 79-94. Russ, Charles V.J. (1982). Studies in historical German phonology: a phonological comparison of MHG and NHG with reference to modern dialects. Berne: Peter Lang. — (1994). The German language today: a linguistic introduction. London: Routledge. Schane, Sanford A. (1973). Generative phonology. Englewood Cliffs, NJ: Prentice-Hall. — (1984). The fundamentals of Particle Phonology. Phonology Yearbook 1. 129-155. Schnitzer, Marc L. (1972). Generative phonology: evidence from aphasia. University Park, PA: The Pennsylvania State University. Scholz, Hans-Joachim (1972). Untersuchungen zur Lautstruktur deutscher Wörter. Munich: Fink. Scobbie, James M. (1991). Towards declarative phonology. In S. Bird (ed.) Declarative perspectives on phonology. Edinburgh: Centre for Cognitive Science. 1-26. Selkirk, Elisabeth 0 . (1978). On prosodie structure and its relation to syntactic structure. In Thorstein Frotheim (ed.) Nordic Prosody II. Trondheim: Tapir. 111-140. — (1982). The syllable. In van der Hulst & Smith (1982b: vol. 2). 337-384. — (1984). On the major class features and syllable theory. In Aronoff & Oehrle (1984) 107-136. Sigurd, Bengt (1955). Rank order of consonants established by distributional criteria. Studia Linguistica 9. 8-20. Slowiaczek, Louisa M. & Daniel A. Dinnsen (1985). On the neutralizing status of Polish word-final devoicing. JPh 13. 325-341. Smith, Norval S.H. (1981). Foley's scales of relative phonological strength. In Goyvaerts (1981b). 587-595. Stampe, David (1969). The acquisition of phonetic representation. CLS 5. 443-454. Stanley, Richard (1967). Redundancy rules in phonology. Lg 43. 393-436. Steriade, Donca (1982). Greek prosodies and the nature of syllabification. PhD dissertation. Cambridge, ΜΑ: ΜΓΓ. Strauss, Steven L. (1982). Lexicalist Phonology of English and German. Dordrecht: Foris. Szigetvári, Péter (1994). The special nature of coronal consonants. University Degree thesis. Budapest: Eötvös Loránd University. Szulc, Aleksander (1987). Historische Phonologie des Deutschen. Tübingen: Niemeyer. Treiman, Rebecca, Jennifer Gross & Annemarie Cwikiel-Glavin (1992). The syllabification of /s/ Clusters in English. JPh 20. 383-402.

272 Tnibetzkoy, Nikolaij S. (1933a). Charakter und Methode der systematischen ponologischen Darstellung einer gegebenen Sprache. Archives néerlandaises de phonétique expérimentale 8-9. 109-113. — (1933b). La phonologie actuelle. Journal de psychologie 30. 227-246. — (1939). Grundzüge der Phonologie. (= Travaux du cercle linguistique de Prague 7). English version, trans. C.A.M. Baltaxe (1969). Principles of Phonology. Berkeley: University of California Press. Ulbrich, Horst (1970). Zur auditiven Interpretation von deutschen /r/-Allophonen. Proceedings of the 6th International Congress of the Phonetic Sciences. 955-958. — (1972). Instrumentalphonetisch-auditive r-Untersuchungen im Deutschen. Berlin: AkademieVerlag. — (1973). Zur Kodifizieiung der R-Aussprache im SIEBS (19. Aufl., 1969). Zeitschrift für Phonetik, Sprachwissenschaft und Kommunikationsforschung 26. 120-133. Vago, Robert M. (1988). Underspecification in the height harmony system of Pasiego. Phonology 5. 343-362. Vater, Heinz (1992). Zum Silben-Nukleus im Deutschen. In Eisenberg et al. (1992). 100-133. Vennemann, Theo (1968). German phonology. PhD dissertation. Los Angeles: UCLA. — (1972). On the theory of Syllabic Phonology. LB 18.1-18. — (1978). Universal syllabic phonology. Theoretical Linguistics 5.175-215. — (1982a). Zur Silbenstruktur der deutschen Standardsprache. In Vennemann (1982b). 261-305. — (ed.) (1982b). Silben, Segmente, Akzente: Referate zur Wort-, Satz- und Versphortologie anläßlich der vierten Jahrestagung der Deutschen Gesellschaft für Sprachwissenschaft, Köln, 2.-4. März 1982. Tübingen: Niemeyer. — (1983). Causality in language change: theories of linguistic preferences as a basis for linguistic explanations. Folia Linguistica Histórica 6. 5-26. — (1988). Preference laws for syllable structure and the explanation of sound change. Berlin: Mouton de Gruyter. Vergnaud, Jean-Roger (1982). On the theoretical bases of phonology. Paper presented at the Vth GLOW Colloquium, Paris. Walsh, Thomas & Frank Parker (1981a). Vowel length and 'voicing' in a following consonant. JPh 9. 305-308. — & — (1981b). Vowel termination as a cue to voicing in post-vocalic stops. JPh 9.105-108. Walshe, M. O'C. (1974). A Middle High German reader. Oxfoid: Oxford University Press. Walther, Markus (1993). Declarative syllabification with applications to German. Edinburgh Working Papers in Cognitive Science 8. 55-79. — & Richard Wiese (1993). Deklarative vs. prozedurale Modellierung von Konsonantenaltemationen im Deutschen. Ms. Düsseldorf: Heinrich-Heine-Universität Düsseldorf. Watson, Ian (1983). Cues to the voicing contrast: a survey. Cambridge Papers in Phonetics and Experimental Linguistics 2. W1-W34. Wells, John C. (1982). Accents of English. 3 vols. Cambridge: Cambridge University Press. Werner, Otmar (1974). Phonemik des Deutschen. Stuttgart: Metzler. Wheeler, Max W. (1979). Phonology of Catalan. Oxfoid: Blackwell. — (1983). Catalan in recent phonological theory. Actes del III Colloqui d'Estudis Catalans a NordAmèrica. Toronto 1982. Montserrat: Publications de l'Abadia. 53-66. — (1986). Catalan sandhi phenomena. In Henning Andersen (ed.) Sandhi phenomena in the languages of Europe. Berlin: Mouton de Gruyter. 475-488. Wiese, Richard (1987). Prosodie conditions on clitics. In Dressier et al. (1987). 331-338. — (1988). Silbische und lexikalische Phonologie: Studien zum Chinesischen und Deutschen. Tübingen: Niemeyer. — (1990). Towards a unification-based phonology. COLING 13. 283-286. — (1991). Was ist extrasilbisch im Deutschen und warum? ZS10.112-133.

273 Williams, Geoff & Wiebke G. Brockhaus (1992). Automatic speech recognition: a principle-based approach. SOAS Working Papers in Linguistics and Phonetics 2. 371-401. Wurzel, Wolfgang Ullrich (1970). Studien zur deutschen Lautstruktur. Berlin: Akademie-Verlag. — (1981a). Die phonologische Komponente. In Heidolph et al. (1981). 145-152. — (1981b). Phonologie: Segmentale Struktur. In Heidolph et al. (1981). 898-990. Yip, Moira (1988). The Obligatory Contour Principle and phonological rules: a loss of identity. LI 19. 65-100. Yoshida, Shohei (1990). A government-based analysis of the 'mora' in Japanese. Phonology 7. 331-351. — (1991). Some aspects of governing relations in Japanese phonology. PhD dissertation. London: SOAS, University of London. Yoshida, Yuko (1990). Government in Tonkawa: vowel elision and cyclicity. SOAS Working Papers in Linguistics and Phonetics 1. 53-69. Yu, Si-Taek (1992a). Silbeninitiale Cluster und Silbifizierung im Deutschen. In Eisenberg et al. (1992). 172-207. — (1992b). Unterspezifikation in der Phonologie des Deutschen. Tübingen: Niemeyer. Zwicky, Arnold M. & Geoffrey Κ. Pullum (1983). Cliticization vs. inflection: English n't. Lg 59. 502-513.