Between Phonology and Phonetics: Polish Voicing 9781614515135, 9781614517146

Citation preview

Eugeniusz Cyran Between Phonology and Phonetics

Studies in Generative Grammar

Edited by Harry van der Hulst Norbert Corver Jan Koster Henk van Riemsdijk

Volume 118

Eugeniusz Cyran

Between Phonology and Phonetics Polish Voicing

DE GRUYTER MOUTON

ISBN 978-1-61451-714-6 e-ISBN 978-1-61451-513-5 ISSN 0167-4331 Library of Congress Cataloging-in-Publication Data A CIP catalog record for this book has been applied for at the Library of Congress. Bibliographic Information published by the Deutsche Nationalbibliothek The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie; detailed bibliographic data are available in the Internet at http://dnb.dnb.de. © 2014 Walter de Gruyter, Inc., Boston/Berlin Printing and binding: CPI buch bücher.de GmbH, Birkach ♾ Printed on acid-free paper Printed in Germany www.degruyter.com

Preface

For decades, the voicing system of Polish has been at the centre of a heated theoretical debate concerning laryngeal phonology. This is due to two factors. Firstly, Polish features a number of phenomena that constitute the core of this debate, such as Final Obstruent Devoicing, Regressive Voice Assimilation, Progressive Voice Assimilation, dialectally distributed external sandhi voicing in pre-sonorant context, and other instances of special behaviour of sonorants with respect to practically all the above mentioned processes, leading to interesting cases of variation. Secondly, Polish data have had very good ambassadors, such as Christina Bethin, Edmund Gussmann, and Jerzy Rubach, whose theoretical work was always at the forefront of the advances in laryngeal phonology, from linear to non-linear accounts, from binary to privative representation of voice, and from the days of the brute force of phonological rules to nonderivational frameworks, seeking to uncover the mechanisms of laryngeal distribution, which would stem directly from the general design of phonological representations or grammar. This book, however, is no less a tribute to the great Polish linguists of the previous century whose work is still highly inspirational. Jan Baudouin de Courtenay draws our attention to the similarity between the external sandhi voicing in Cracow-Poznań Polish and similar facts in Sanskrit, thus broadening horizons and prompting a universalist, or at least a comparative perspective on voicing issues. Tytus Benni’s generalizations attempting to incorporate Progressive Voice Assimilation into a uniform system of assimilations, rather than seeing them as an exception to Regressive Voice Assimilation, are echoed in most modern analyses. Kazimierz Nitsch provided many insights into the nature of the external sandhi voicing in Cracow-Poznań Polish and Progressive Voice Assimilation. Finally, Zdzisław Stieber proposed a still valid structural and systemic explanation of Progressive Voice Assimilation, while his inspirational comments about the relationship between phonology and phonetics contributed greatly to the model of Sound System, which underlies the analyses of Polish voicing in this book. The database used in this book is a combination of well-established facts presented in previous modern and earlier analyses, with recent experimental results, which point to greater variation in all the processes mentioned above than formal approaches assumed, or were ready to admit. Gradient phenomena

vi | Preface

and variation are difficult to express in formal descriptions, especially if they do not envisage a clear division between what is phonological and what is not phonological in the phenomena under study. On the other hand, when phonology is relieved of most of the explanatory responsibility, the implementational and indeed purely phonetic factors, as well as their relationship with phonology proper, needs to be spelled out, especially in models which eschew phonetic grounding in phonology. This book attempts to do that, and more. The model presented here is one in which phonology, phonetic interpretation and phonetics find a home, or rather their respective homes, in fact. Paradoxically, by separating these three levels of description, we wish to integrate the disparate threads of modern research of sound patterns into one sound system. For example, we take seriously the major results of laboratory phonology research, which gives us insight into how production and perception shape sound patterns, as well as seemingly competing theories of phonological organization. An example of the latter, would be the inspiring hypothesis of Licensing by Cue (Steriade 1999), which, although couched within Optimality Theory, does not contradict our own proposals stemming from a representation-based theory of Government in phonology. Since OT attempts to model the grammar with a view to arriving at the end point of ‘derivation’, that is, the optimal surface forms, it is by definition closer to implementational and phonetic aspects of sounds systems. The theories of representations, then, are not, by definition incompatible with theories of constraint interaction. The question is how phonetic and implementational patterns can be incorporated into the computational module of phonology, and how much of this type of information fits in. At any rate, as research into laryngeal phonology progresses on various fronts, it becomes more and more obvious that a large portion of the phenomena in question have phonetic, or implementational conditioning, thus limiting the role of phonology even further. One of the aims of this study is to show the nature of what is left, and must remain. The phonological model used in this study is a minimalist representational approach stemming from the research programme, which is generally referred to as Government Phonology (Charette 1991; Harris 1990, 1994; Kaye 1990, 1995; Kaye, Lowenstamm and Vergnaud 1985, 1990). The exact version of that theory will be that proposed in Cyran (2010), which follows the view that phonological representation is formally organized by a series of CVs, as well as lateral relations of Government and Licensing (Lowenstamm 1996; Polgárdi 1998; Rowicka 1999; Scheer 2004). The representation of melody in this book follows the main tenets of the Element Theory (Harris 1990, 1994; Harris and Lindsay 1993, 1995), of which the most important ones are as follows. Melodic representations are privative. Non-specified members of oppositions do not receive further specifi-

Preface | vii

cation during derivation. They are interpreted phonetically as such. With respect to voicing oppositions, Laryngeal Realism (Harris 1994; Honeybone 2002, 2005; Iverson and Salmons 2003a) will be used as a starting point of reference. The discussion here takes a slightly different course than usual, in that the central aspects of Polish voicing such as Final Obstruent Devoicing (FOD) and Regressive Voice Assimilation (RVA) are given less focus than Cracow-Poznań sandhi voicing (CPV) and Progressive Voice Assimilation (PVA). In this sense, we adhere to the early descriptions of Polish in which these ‘fringe’ phenomena drew more energy and analytic effort because they are more difficult to deal with. The usual course of action in more recent analyses has been to establish a model which accounts for the major phenomena, that is, FOD and RVA, and then try to demonstrate, with varying degrees of success, how the model applies to PVA and CPV. The net result has almost invariably been a postulation of a separate rule, or particular extrinsic rule ordering which contradicted the original intuition that all these phenomena are intrinsically connected. The book is organized as follows. Chapter 1 contains preliminary information concerning the Polish voicing data and discusses the main theoretical issues connected with laryngeal phonology. The model of Laryngeal Realism is introduced and applied to the data with a view to demonstrating its failure with respect to CPV. Chapter 2 introduces a modification of the above mentioned model, which will be referred to as Laryngeal Relativism. The new model is placed in a broader perspective of a Laryngeal System, which, like any Sound System based on arbitrary relations between phonological and phonetic categories, is composed of three levels of description: phonological, implementational (phonetic interpretation) and phonetic. The chapter contains definitions of various types of voicing found in natural languages and argues for a new type: enhanced passive voicing, which, it is claimed, occurs in Cracow-Poznań Polish. Chapter 3 provides a new analysis of CPV, in which no new rule or rule ordering is necessary. The phenomenon is claimed to be purely phonetic and interpretational. The analysis, however, works under one condition: the phonological representation, and therefore the phonetic interpretation conventions, are the opposite of what is found in Warsaw Polish. In other words, two different systems generate almost identical sound patterns, with the same series of voiced and voiceless unaspirated obstruents and the same phenomena such as FOD and RVA, except that in external sandhi the interpretation of the neutral obstruents must be voiced in Cracow-Poznań and voiceless in Warsaw Polish for systemic reasons. Chapter 4 contains a rather lengthy discussion of Progressive Voice Assimilation, in which an attempt is made to explain the phenomenon by devoting close attention to the distinction between phonological, interpretational and phonetic conditioning. This chapter enforces a reinterpretation of

viii | Preface

such phenomena as obstruency, obstruentization and obstruent devoicing, as well as the sometimes automatic connection which is made between friction and obstruency. With the new results it is possible to demonstrate that PVA is not an exception to RVA because it has a completely unrelated conditioning. The concluding Chapter 5 is a collection of issues that either have been left unexplored in the previous discussion, or had to be left to this point in the book for presentational reasons. The reader will find a detailed proposal of a uniform formal source of laryngeal licensing, which is at the same time amenable to micro-shifts leading to the observable variation in data. A section is devoted to the behaviour of Polish prepositions and prefixes. We also return to the problem of sonorant behaviour with respect to spreading processes, and the relationship between phonology and phonetics. One of the consequences of the model presented in this work is that subsegmental primes – here elements – can be viewed as devoid of phonetic substance. The chapter ends with a brief reference to two languages in which the overall laryngeal system seems to resemble that of Cracow-Poznań Polish. We look at possible analytical consequences of this. I wish to express my gratitude to the following friends and colleagues for their generous assistance and comments at various stages in the writing of this book. First and foremost, many thanks are due to my phonology teacher and friend, the late Edmund Gussmann. He did not have a chance to see any of this book, but I am sure his comments would have changed it a lot. I am also extremely grateful to Jonathan Kaye and Tobias Scheer, who spent time talking to me about phonology including laryngeal phonology. I am not sure if we fully agree, but their encouragement allowed me to press on with my ideas. Thanks are also due to Sławek Zdziebko, who looked at the manuscript in great detail, to Robert Looby for proofreading the text, and to Katarzyna Bednarska, Ewa Pająk and Krzysztof Jaskuła for providing other comments. Last but not least, I would like to thank the anonymous reviewer and Harry van der Hulst for all their help and comments. I take full responsibility for the errors that remain. Last but not least, I would like to thank my wife Marta and daughter Jadzia for their patience and support. This book is dedicated to them.

Contents Preface | v Chapter 1: Preliminaries | 1 1 Introduction | 1 2 The basic Polish voicing facts | 1 3 Representing voice | 5 4 Laryngeal Realism | 8 5 Laryngeal Realism and Polish voicing | 11 6 The problem of Cracow-Poznań sandhi voicing | 17 Chapter 2: Sound system, phonology and phonetics | 21 1 Introduction | 21 2 Types of voicing | 22 2.1 Spontaneous voicing | 22 2.2 Active voicing | 25 2.3 Passive voicing | 26 3 Sound system | 29 3.1 Phonological aspects | 30 3.2 Phonetic categories | 32 3.3 Phonetic interpretation | 34 4 Laryngeal Relativism | 37 5 Some consequences of Laryngeal Relativism | 40 Chapter 3: Cracow-Poznań sandhi voicing | 43 1 Introduction | 43 2 Laryngeal Relativism and Polish voicing | 44 2.1 Final Obstruent Devoicing | 47 2.2 Regressive Voice Assimilation | 52 3 Cracow-Poznań sandhi voicing | 56 3.1 Bethin (1984, 1992) | 59 3.2 Gussmann (1992) | 64 3.3 Rubach (1996) | 66 3.4 CP sandhi voicing and Laryngeal Relativism | 68 3.5 Laryngeal Relativism and Licensing-by-Cue | 73 4 Conclusions | 76

x | Contents

Chapter 4: Progressive voice assimilation in Polish | 79 1 Introduction | 79 1.1 Benni’s generalization | 83 1.2 Is PVA dead or alive? | 86 1.3 A functional explanation of progressive devoicing | 89 2 Modern derivational and non-derivational accounts of PVA | 90 2.1 Preliminaries | 90 2.2 Gussmann (1992) | 93 2.3 Bethin (1992) | 99 2.4 Rubach (1996) | 103 2.5 Gussmann (2007) | 107 3 PVA and Laryngeal Relativism | 113 3.1 Preliminaries | 113 3.2 Laryngeal Relativism | 113 3.3 Obstruentization and PVA | 118 3.4 The ‘voiced obstruent’ hypothesis in CP and WP | 124 3.5 The new (old) analysis of PVA | 127 3.6 Post-alveolar fricatives in PVA context | 132 4 Conclusions | 138 Chapter 5: Further issues | 141 1 Introduction | 141 2 Laryngeal licensing – elements in phonological contexts | 142 3 Prepositions and voicing | 159 4 The behaviour of sonorant consonants | 166 4.1 ...C# vs. ...CS# | 168 4.2 ...CS#C... vs. ...C#SC... | 177 Phonology and phonetics in Laryngeal Relativism | 193 5 5.1 Substance-free elements? | 198 6 Laryngeal Relativism and other languages | 201 6.1 Durham English | 201 6.2 Dutch | 205 7 Conclusions | 200 Conclusion | 213 References | 217 Subject index | 229 Index of Polish words | 231

Chapter 1 Preliminaries 1 Introduction In this chapter, we look first at the least controversial and best known voicing facts from Polish, which include the type of laryngeal contrast utilized in this system as well as the two core phenomena, namely Final Obstruent Devoicing (FOD) and Regressive Voice Assimilation (RVA) in obstruent clusters. This brief survey will not include details connected with the behaviour of sonorants, prefixes and prepositions, or the so called Progressive Voice Assimilation (PVA). These issues will receive attention in separate chapters once the new model of description is introduced in Chapter 2. The data discussed below will serve the purpose of allowing us to set the scene in terms of basic facts and theoretical choices that need to be made in order to grasp the entire voicing complex of Polish. As a starting point to a more detailed theoretical discussion we introduce the approach called Laryngeal Realism and apply it to the data with a view to demonstrating better the theoretical and analytical problems that Polish data pose especially on privative models of laryngeal representation. The chapter ends with an introduction of one of the most complicated problems of the Polish voicing complex, namely, the Cracow-Poznań sandhi voicing before sonorants. It will be shown that Laryngeal Realism is unable to provide an explanatory account of this phenomenon.

2 The basic Polish voicing facts The basic facts of the Polish voicing system are well known,1 both from the traditional phonetic studies (e.g. Benni 1964; Wierzchowska 1971), but especially from a number of generative and post-generative analyses (e.g. Bethin 1984, 1992; Gussmann 1992, 2007; Rubach 1996, 2008; Michalski 2009). Polish has a

|| 1 There are a number of experimental studies rejecting or supporting such ‘established’ assumptions as word-final neutralization or sonorant transparency and opacity (e.g., Slowiaczek and Dinnsen 1985; Jassem and Richter 1989; Strycharczuk 2012b).

2 | Preliminaries

two-way voicing contrast among obstruents, which is typically depicted as fully voiced versus voiceless unaspirated. The system has also been described as fortes versus lenes (Benni 1964: 19, Wierzchowska 1971: 149), but unlike in German or English in which the fortes series is accompanied by aspiration in stops, as well as a number of other cues, the distinction along the articulatory strength parameter does not seem to have any consequences in Polish.2 The relatively stronger constriction in the voiceless unaspirated series is assumed to be a natural phonetic consequence of the lack of voicing.3 The data in (1) illustrate some minimal pairs with the voiced / voiceless contrast. (1)

pić bić płotem błotem rysa ryza oknie ognie

[pjit] [bjit] [pwtm] [bwtm] [rsa] [rza] [k] [g]

‘to drink’ ‘to hit’ ‘fence, instr.’ ‘mud, instr.’ ‘scratch’ ‘ream’ ‘window, loc.’ ‘fire, pl.’

The contrast is neutralized in two contexts: before an obstruent and wordfinally. Examples of FOD are given below and include devoicing of obstruent clusters, as well as obstruents followed by sonorants.4 (2) Final Obstruent Devoicing waga stogu żaba koza

[vaga] [stgu] [aba] [kza]

/ / / /

wag stóg żab kóz

[vak] [stuk] [ap] [kus]

‘scale, nom.sg./gen.pl.’ ‘haystack, gen.sg./nom.sg.’ ‘frog, nom.sg./gen.pl.’ ‘goat, nom.sg./gen.pl.’

|| 2 Dłuska (1950: 79) describes the difference between the articulations of voiceless stops in Romance and Slavic on the one hand, and Germanic on the other, as having a closed and raised glottis in the former, and open in the latter group. 3 It should be mentioned, however, that the articulation of the voiced / voiceless opposition in Polish exhibits a number of articulatory parameters which are typically found in the fortis / lenis one. With respect to [t/d], Wierzchowska (1971: 154) enumerates the following features of [t]: stronger burst, stronger laminal articulation, raised larynx, and longer duration of closure. 4 Recently, researchers are more prone to admit the variability in the operation of FOD in forms such as kadr [katr] ‘personnel, gen.pl.’. This issue will be returned to in Chapter 5.

The basic Polish voicing facts | 3

gwiżdżę [gvjid] / gwiżdż [gvji t ] ‘I whistle/whistle, imp.’ mózgu [muzgu] / mózg [musk] ‘brain, gen.sg./nom.sg.’ gwiazda [gvjazda] / gwiazd [gvjast] ‘star, nom.sg./gen.pl.’ blizna dobro kadra

[blizna] / blizn [dbr] / dóbr [kadra] / kadr

[blisn] [dupr] [katr]

‘scar, nom.sg./ gen.pl.’ ‘goodness, nom.sg./gen.pl.’ ‘personnel, nom.sg./gen.pl.’

Word-final devoicing results in surface ambiguities. For example, a surface form [stuk] may have two lexical sources, that is, stóg ‘haystack’ with final devoicing, or stuk ‘a knock’, with a lexically voiceless obstruent. As observed in (2) above, obstruent clusters which were lexically voiced, e.g., gwiazda [gvjazda] ‘star, nom.sg.’, become voiceless as a whole in gwiazd [gvjast] ‘star, gen.pl.’. This brings us to the other context in which voicing cannot be used contrastively, and illustrates the general phonotactic pattern found in Polish clusters of obstruents. Either they statically agree in voicing as in (3a), in which case we are dealing with a Voice Agreement, or they become uniform in voicing by assuming the value of the right-hand obstruent. This phenomenon is called Regressive Voice Assimilation and occurs inside words (3b), across morpheme boundaries (3c), and across word boundaries (3d). (3)

Static voice agreement a. kto gdy brzask ksiądz krtań drgać

[kt] [gd] [bask] [knts] [krta] [drgat]

‘who’ ‘when’ ‘dawn’ ‘priest’ ‘larynx’ ‘tremble’

*kd, *gt *kd, *gt *b , *p, *zk, *sg *k , *g *krd, *grt *drk, *trg

Dynamic voice agreement (assimilation) b. dech [dx] / tchu [txu] ‘breath, nom.sg./gen.sg.’ wieś [vj] / wsi [fi] ‘village, nom.sg./gen.sg.’ wesz [v ] / wszy [f ] ‘louse, nom.sg./gen.sg.’ c. prosić [prit] / prośba [pr ba] ‘to ask/a request’ ryza [rza] / ryzka [rska] ‘ream/dim.’ mędrek [mndrk] / mędrka [mntrka] ‘smart-aleck/gen.sg.’ d. kwiat begonii [kfjad bgji] ‘begonia flower’ litr bimbru [ljidr bjimbru] ‘a litre of moonshine’

4 | Preliminaries sad śliwkowy [sat ljifkv] szyb kopalni [ p kpali]

‘plum orchard’ ‘mine shaft’

The data in (3) require some comments. Firstly, the stem-internal alternations of voice in dech / tchu (3b) are extremely rare in Polish, which is why the forms in (3b) are quoted in almost every analysis of voicing facts in Polish. One reason for the special status of such forms is the fact that, two obstruents of opposite voice value must additionally be separated by a vowel that alternates with zero. Thus, when the vowel disappears, as in tchu, wsi, or wszy, the two obstruents become adjacent and their opposite voice value needs to be brought in line with the Voice Agreement condition. Interestingly, there seems to be no example of the reverse type of assimilation, that is, to a voiced obstruent.5 Similarly restricted are the assimilations across morpheme boundaries, as in (3c). This is due to the existing list of suffixes in Polish. Here, the assimilation to a voiced obstruent is basically limited to the fairly unproductive nominalizing suffix –ba, e.g., prosić [prit] ‘to ask’ vs. prośba [pr ba] ‘request’, and the context before the clitic –by, rósłby [ruz(w)b] ‘would grow’. On the other hand, the assimilation to a voiceless obstruent is generally limited to the suffix −ek/−ka. There are two data sets which to some extent compensate for this deficiency. Firstly, there is a relatively high incidence of forms involving the static agreement in sequences of two obstruents, as in (3a). And secondly, there are numerous cases of assimilation to a voiced obstruent across a word boundary, as illustrated in (3d), e.g., kwiat begonii [kfjad bgji] ‘begonia flower’. Here, unlike with the root-internal and word-internal situation, the increased incidence of such cases is directly related to the number of words that begin with voiced obstruents in the Polish lexicon. The data in (3d) include instances of assimilation across word boundaries which appear to be cases of voicelessness spreading, e.g., szyb kopalni [ p kpali] ‘mine shaft’. Since the lexical voiced obstruent in the word szyb [ p] ‘shaft’ is devoiced word-finally anyway, it remains to be seen if we are dealing with assimilation here. A better example of regressive assimilation to a voiceless obstruent involves prefixes or prepositions followed by words beginning with a voiceless obstruent, e.g., pod kopalnią [pt kpalw] ‘under the mine’. Polish prefixes and prepositions do not undergo FOD if not used in isolation, which || 5 This gap is to a great extent accidental as it stems directly from the historical development of yer vowels. However, it has some consequences, which will be considered in more detail in connection with the intriguing phenomenon of so called progressive devoicing in, e.g., krewny [krvn] ‘a relative’ vs. krwi [krfji] ‘blood, gen.sg.’. There are reasons to assume that the labiodental fricative is lexically voiced, which should produce *[grvji], parallel to drgać in (3a).

Representing voice | 5

can be seen in the forms where the following word begins with a vowel, e.g., pod oknem [pd knm] ‘under the window’, or a sonorant consonant, e.g., pod mostem [pd mstm] ‘under a bridge’ (Booij and Rubach 1987). Thus, clearly, in pod kopalnią, the devoicing of the preposition-final obstruent is due to the voiceless object at the beginning of the following noun. Thus, we can conclude that at least in surface terms, we are dealing with a symmetrical phenomenon of voice agreement in obstruent clusters, that is, both to a voiced and to a voiceless trigger, a point that will be returned to in the following sections. Given the basic data above, the distribution of the voicing distinction, if expressed by the ability to maintain the laryngeal distinction (Lar), is easily captured in descriptive terms. The voice contrast is maintained before vowels ‘_V’ and before sonorant consonants followed by vowels ‘_SV’. The two contexts can be schematically merged into one ‘_(S)V’ as shown in (4a). The contrast is neutralized word-finally, whether the sonorant is present or not ‘_(S)#’ (4b),6 and before other obstruents (4c). (4) a.

b.

... C (S) V... | Lar

c. ... C(S) # | Lar

... C (S) C... | Lar

C = obstruent, (S) = optional sonorant, Lar = laryngeal specification, V = vowel Having seen the basic data concerning the distribution of the voice contrasts in Polish, we now turn to some theoretical preliminaries leading to the specific choice of a representational model that will be a starting point to our own proposal in Chapter 2.

3 Representing voice For any phonological model which recognizes a distinction between phonology and phonetics, and that concerns most models, the following simplified scheme (paradigm) applies. However, the difference between individual approaches may produce quite a range of possibilities: from full incorporation of phonetic grounding into phonology to substance free phonology.

|| 6 Some variation is present in forms like kadr [katr ~ kadr] ‘personnel, gen.pl.’. See Chapter 5.

6 | Preliminaries

We limit the discussion of the phonological representation only to the relevant melodic aspect, that is, laryngeal specification. There are a few fundamental choices that may be made here, and the particular options greatly influence the rest of the paradigm. (5) Phonological representation

Computation (Rules)

/.../

Phonetic form [...]

One such decision is whether the distinctive features are binary, for example, [&voice], [&spread glottis], [&constricted glottis], [&stiff vocal folds], [&slack vocal folds], or whether the laryngeal contrasts should be expressed privatively. In the latter case, only the marked segments contain a particular feature, for example [voice], and the unmarked segments do not have any laryngeal specification underlyingly.7 One of the consequences of the above decision is the way basic phonological phenomena related to voice will be described. For example, Final Obstruent Devoicing (FOD) in Polish as in żaba / żab [aba ~ ap] ‘frog, nom.sg./gen.pl.’, will receive quite disparate analyses. In a binary model, the phonological rule may take the following form (e.g., Hayes 1984: 319). (6) Final Devoicing C → [–voice] / __ # Problems with rules of this type concern their purely descriptive adequacy and arbitrariness. If both [+voice] and [–voice] are available in computation, then such rules as in (6) do not explain, for example, why an obstruent is devoiced and not voiced in that environment. This is because the representational symmetry predicts a computational one. The analysis is partly improved by breaking the rule in (6) down into two separate ones, as in Rubach (1996: 77).

|| 7 Given that we are mainly concerned with the laryngeal system of Polish, we limit the discussion to a two-way contrast, in which one series of consonants is marked and the other one remains unmarked.

Representing voice | 7

(7)

a. Final Devoicing R L

[–sonor] / __ )PW

b. Voice Default R | [–sonor] →

R L [–sonor] | [–voice]

One rule delinks the laryngeal node of the obstruent in word-final position, regardless of whether it had [+voice] or [–voice] (7a). Thus, technically speaking, it is a delaryngealization and not merely a devoicing rule. The other rule is a universal default, which reintroduces the laryngeal node on unspecified obstruents at a later stage in the derivation and fills it in with [–voice]. Here, one of the results of the improvement is the introduction of rule ordering, where (7a) feeds (7b). The other consequence is that the derivation has intermediate stages at which, next to the phonetically observable voiced and voiceless obstruents, a third object is created. It is an unspecified obstruent, which cannot be interpreted phonetically unless it receives some specification in further derivation. The universal tendency for obstruents to be voiceless rather than voiced is builtin in the Voice Default but we are still dealing with mere descriptive adequacy because rule (7b) only formalizes the observation and does not explain it. Privative models assume that in a two-way contrast system like Polish only one series of obstruents carries laryngeal specification. Typically, it is assumed that the voiced series contains the feature [voice]. The models differ, however, with respect to how the unmarked series is treated in further derivation. In the approaches in which phonological derivation is meant to bring the phonological representation closer to the stage at which it becomes phonetically interpretable – systematic phonetic level – the underlyingly unmarked segments receive their respective laryngeal features by means of various defaults. Universally, sonorant consonants and vowels become [+voice], while obstruents become [–voice]. We find nothing wrong with defaults as long as they are mere phonetic interpretation statements, and the filled values do not participate in any phonological computation. Unfortunately, they do in a number of accounts of the Polish voicing system (e.g., Bethin 1992). In this sense, it is difficult to maintain that we are indeed dealing with a privative system. A more radical proposal is made, for example, within the Element Theory of Government Phonology (GP). Instead of underspecification, GP talks about nonspecification of the unmarked objects and assumes that the segments receive phonetic interpretation without any feature filling devices (Kaye, Lowenstamm, and Vergnaud 1990; Harris 1990, 1994). Phonological representations in this

8 | Preliminaries

model are phonetically interpretable at any stage of the derivation. This does not mean that there are no phonological processes or computation. To the contrary. However, the computation does not bring the phonological representation any closer to the level at which they would be more pronounceable. The representations are pronounceable both before and after phonological processing, and the systematic phonetic level of representation is dispensed with (Harris and Lindsey 1993, 1995). As a starting point, we assume the position called Laryngeal Realism 8 (Harris 1994; Honeybone 2002, 2005), which is a privative approach with nonspecification rather than underspecification. We review this approach below concentrating on how the privative view accounts for basic phenomena such as FOD and Voice Assimilation (VA), as well as on the question of what is a possible phonological rule (computation).

4 Laryngeal Realism Laryngeal Realism (Honeybone 2002, 2005) is an approach to laryngeal phonology whose main assumption is the privativity of the representation of voice coupled with non-specification and direct phonetic interpretation of the unmarked segments familiar from the Element Theory of Government Phonology (e.g., Kaye, Lowenstamm and Vergnaud 1985, 1990; Harris 1990, 1994). It is based on two important observations from the realm of universal phonetics and translates them into phonological systems. The first observation is that there are three major phonetic categories which are regularly utilized by languages to express laryngeal contrasts.9 These categories can be illustrated along the so called VOT continuum (Lisker and Abramson 1964; Lieberman 1970; Keating 1984), i.e., i) long lead (negative VOT, which is found in fully voiced stops), ii) short lag (voiceless unaspirated stops), and iii) long lag (voiceless aspirated stops). The symbols used below are taken from Honeybone (2002).

|| 8 We owe the term to Honeybone (2002) but in spirit the proposal is not different from Harris (1994). For this reason we refer to both authors here. 9 We exclude systems with glottalization in this discussion. It is possible within Element Theory (e.g., Harris 1990, 1994) to represent this property by means of the occlusion element { } rather than a special laryngeal category, e.g., [constricted glottis].

Laryngeal Realism | 9

(8)

fully voiced

voiceless unaspirated

voiceless aspirated

b d g

po to ko

ph th kh

The other observation, this time concerning mostly European languages, is that among languages possessing a two-way laryngeal contrast, one can observe two groups of languages according to the different kinds of laryngeal distinctions they use. The first group is said to be represented by most Germanic languages and is characterized by a fortis / lenis distinction, which translates into voiceless aspirated vs. voiceless unaspirated in surface terms. The other group, represented by Romance and Slavic languages, has an opposition between fully voiced and voiceless unaspirated stops. Additionally, the two groups tend to exhibit different patterns of assimilation. Languages with full voicing seem to spread that property in assimilation processes, e.g., Polish rendition of football as [fudbl] vs. English [f tbl]. Laryngeal Realism takes these facts as evidence that the two groups of languages cannot have the same phonological representations of the contrast that would be based on the feature [voice], and proposes two different types of representation.10 (9)

Harris Honeybone

Romance and Slavic

Germanic

fully voiced

voiceless unaspirated

voiceless aspirated

b

po

ph

{L} [voice]

− −

{H} [spread]

The Germanic languages are viewed as ‘aspiration’ systems in which the marked feature, or in fact element, is high tone {H} for Harris (1994), and

|| 10 One should note the proposal of Keating (1984) who uses [–voice] and [+voice] as phonological categories and derives surface phonetic categories corresponding to those in (8).

10 | Preliminaries [spread] for Honeybone (2002).11 The Romance and Slavic languages are assumed to base the opposition on [voice], which translates into the low tone element {L} for Harris and the element |voice| for Honeybone. Note in the graph above that in both systems the voiceless unaspirated objects symbolized with /po/ are neutral, i.e., non-specified.Given the presence of a non-specified object in both systems, Harris (1994: 135) derives a more general typology of phonation systems based on this fact and on the use of two laryngeal elements, which he has at his disposal. Thus, in a pre-contrast system – which only has one series of obstruents – no laryngeal specification is employed at all. An unmarked series is also present in two-, three-, or even four-way contrast languages. The growing complexity of the systems is directly correlated with the number of laryngeal elements used and whether they themselves may combine, as in Gujarati. (10)

representation Hawaiian Polish English Thai Gujarati

{L} {L} {L}

{_} {_} {_} {H} {_} {H} {_} {H} {L-H}

examples p b p b ph b p ph b p ph b 

Returning now to the non-specified obstruents, it should be remembered that one of the distinguishing factors among the privative models is the treatment of the unmarked series. Thus, in some traditions it is lexically underspecified but receives phonetic content (features) in derivation (e.g., Iverson and Salmons 1995, 2003b; Bethin 1992). In the Laryngeal Realism tradition (Harris 1994, 2009; Honeybone 2002, 2005), on the other hand, no systematic level of phonetic representation is postulated at which more concrete representations are derived. The unmarked segment is non-specified and it is directly interpretable on a language specific basis. Nevertheless, the two privative traditions are similar in exhibiting a strong phonetic bias with respect to the analytical decision as to what actual phonological representation stands behind the observable surface facts. For example, the presence of full voicing is taken to be the indication of the presence of the element [voice], or {L}, while aspiration leads to the postulation of [spread], or {H}.

|| 11 Honeybone uses a different notation, namely, |spread| to express the fact that it is not a feature but an independently interpretable element. We keep the square brackets here for expository reasons.

Laryngeal Realism and Polish voicing | 11

Typically, the authors do correlate the respective representations also with phonological phenomena such as presence or absence of a particular type of assimilation. However, while phonetically correct, it is not impossible to provide alternative and equally valid analyses, as we will see below. It is often argued within privative frameworks that assimilation is possible only if the active feature / element is spread. If we talk about phonological assimilation only, this is correct. However, the fact that assimilation as a phonetic fact can be symmetrical, that is, both to voiced and to voiceless segments in e.g., Polish, suggests that either privativity is on the wrong track, or that there is also a phonetic or interpretational assimilation, which has little to do with active phonological categories. We will see examples of such analyses presently. Not only assimilations can be given alternative analyses. Harris (2009) convincingly argues that not every case of final devoicing is in fact phonological delaryngealization. All the above points strongly suggest that an analysis of a particular voicing system should do much more than look at the acoustics and the presence of assimilation and devoicing. We will return to this point in our further discussion. If the division into H- and L-systems could be done so easily on the basis of observable VOT values, half of the analytical job would be done by looking at the spectrograms. We would also expect little variation between the actual phonetic details in individual systems. On the other hand, not only do such details differ, but also they are difficult to understand under the view that VOT lead languages are L-systems, and VOT lag languages are H-systems. At least, as will be demonstrated presently, it would not be so difficult to understand the laryngeal system of Polish. Let us look at how the basic voicing facts from Polish are handled by Laryngeal Realism with a view to demonstrating how this model works on actual data and defining where its limitations lie.

5 Laryngeal Realism and Polish voicing Let us begin the discussion by clarifying the notation to be used as shorthand for full representations. Following the arguments laid out in Harris (1994), and especially in Honeybone (2002: 141-142), we will use two types of symbols for Polish obstruents depending on their lexical laryngeal specification. It will be recalled from the previous section that Honeybone uses a system of three symbols /po, ph, b/ to refer to neutral, voiceless aspirated and fully voiced stops respectively in the two groups of languages, that is, Germanic and Slavic. We adapt this notational system a little to fit the Polish system, and especially to facilitate the modification of Laryngeal Realism, which will be proposed in the

12 | Preliminaries

following chapter. In short, when referring to obstruents in some abstraction the symbols /Co, CH, CL/ will be used, corresponding to Honeybone’s /po, ph, b/. On the other hand, when referring to concrete examples, we will use, e.g., /po/ to refer to the lexically neutral (voiceless unaspirated) [p], as in pić /pjoit/ > [pjit] ‘to drink’, and /bL/ to refer to a lexically voiced [b], as in być /bLt/ > [bt] ‘to be’.12 Thus, in discussing phonological representations, the symbols will always be accompanied by some diacritic, that is, /po/ or /bL/. Understandably, in the actual phonological representation, the zero diacritic does not correspond to any reality. Throughout this book, the symbol ‘>’ will be used to mean ‘phonetically interpreted as’. Phonological processes, or any change, on the other hand, will be marked with ‘→’. As mentioned in the previous section, for Laryngeal Realism, a language like Polish, which has a phonetic contrast between fully voiced and voiceless unaspirated obstruents, and in which assimilation to a voiced obstruent is observed, should be represented by the presence of the element {L} in the voiced series, while the voiceless unaspirated obstruents are lexically non-specified. Recall that in the non-neutralizing context, that is, ‘_(S)V’ (4a), these values remain unchanged by any phonological processing and are phonetically interpreted as voiced and voiceless unaspirated, e.g., /tom/ > [tm] tom ‘volume’ vs. /dLm/ > [dm] dom ‘house’.13 More examples are given below. (11) ...C(S)V... C retains its lexical laryngeal specification #CoV #CLV #CoSV #CLSV VCoV VCLV VCoSV VCLSV

/pjoito/ > jL o /b it / > /powtom/ > /bLwtom/ > /rsoa/ > /rzLa/ > /ko/ > /gL/ >

[pjit] [bjit] [pwtm] [bwtm] [rsa] [rza] [k] [g]

pić ‘to drink’ bić ‘to hit’ płotem ‘fence, instr.’ błotem ‘mud, instr.’ rysa ‘scratch’ ryza ‘ream’ oknie ‘window, loc.’ ognie ‘fire, pl.’

The actual theory internal explanation for the survival of the laryngeal specification in ‘_(S)V’ will be provided in Chapter 5. As a preview of that analysis we may use the concept of licensing, which is compatible with the mechanisms || 12 For reasons of exposition, we choose to use Harris’s element symbols, that is {L}, and later in the discussion also {H}, corresponding to Honeybone’s [voice] and [spread]. 13 Technically speaking, sonorants should also bear the superscript ‘o’, as they are assumed to be devoid of any laryngeal specification.

Laryngeal Realism and Polish voicing | 13

employed in the Element Theory in Government Phonology (e.g., Gussmann 2007; Harris 1990, 1994). In syllabically oriented analyses of Polish voicing (e.g., Bethin 1984, 1992; Gussmann 1992), this context corresponds to the onset position, or the head of a complex onset. And conversely, the neutralization / delaryngealization context has been identified as the coda. A delaryngealized (neutralized) lexical /bL/ will be referred to as /bo/, where the voiced symbol is a mere transcriptional trace of the object we are dealing with lexically. Its phonetic interpretation, however, will be identical to that of /po/, e.g., /abLa/ > [aba] żaba ‘frog’ vs. /abL/→/abo/ > [ap] żab ‘frog, gen.pl.’. Note, that a phonological process has applied in the genitive plural form, by which /bL/ becomes /bo/. Since the phonetic interpretation of a neutral obstruent in Polish is that of voiceless unaspirated object, /bo/ will have exactly the same interpretation as a lexical /po/, namely, [p]. Thus, the phonetic interpretation takes into account only the superscripted value in our phonological transcription. Phonologically speaking, a delaryngealized /bo/ and a lexical /po/ are identical objects. We saw in (4b,c) that the laryngeal contrast in Polish is suspended wordfinally and before another obstruent. The former situation leads to FOD, while the latter effects Voice Assimilation. FOD results in surface ambiguities. For example, a surface form [stuk] may have two lexical sources. It is either stóg [stuk] ‘haystack’ with final devoicing (cf., stogi [sti] ‘haystack, nom.pl.’), as in (12a), or stuk [stuk] ‘a knock’ (cf., stukanie [stuka] ‘knocking’) with a lexically voiceless obstruent (12b). Below, only the relevant consonants are given the superscripted diacritics. (12)

Lexical representation

Phonological representation

Phonetic interpretation

L-delinking (FOD) L

a. /stug /

→

/stugo/

>

[stuk]

b. /stuko/

=

/stuko/

>

[stuk]

Since the element {L} does not survive in word-final position, it is delinked in stóg (12a). This derivation illustrates another aspect of GP. The lexical representation, which is in fact a fully interpretable phonological representation, may be subject to phonological computation. Here, we are dealing with the delinking or deletion of the laryngeal category in a hostile environment ‘_#’.14 What remains in || 14 A precise explanation how this environment works will be given in Chapter 5.

14 | Preliminaries

that position is a delaryngealized object which is identical to the non-specified series in that system, and must receive the same phonetic realization, that is, voiceless unaspirated. It is important to note that a potential absence of Ldelinking would still produce a fully interpretable representation. The obstruent would then have to be pronounced as voiced. On the other hand, in (12b), the lexical representation is not subject to any processing, and its interpretation is obvious. To repeat, the difference between /bo/ and /po/ is not phonological. Phonologically they are the same, except that /bo/ is an output of delaryngealization. Turning now to the phenomenon of Regressive Voice Assimilation (RVA), we recall that, unlike FOD, it appears to be symmetrical in Polish. That is, we witness assimilations to a voiced and to a voiceless obstruent. In systems using binary features (e.g., Gussmann 1992; Rubach 1996), the analysis is straightforward. First, the left-hand obstruent must lose its laryngeal specification, and then the [+voice] or [–voice] spreads leftwards from the following one. Let us illustrate this point with the rules proposed in Rubach (1996: 78).15 (13)

a. Obstruent delinking R [–sonor]

R L L

[–sonor]

b. Spread R

R L

[–sonor]

Rule (13a) consists in the delinking of the laryngeal node in an obstruent if the node is adjacent to the laryngeal node of the following obstruent. The laryngeal nodes contain either [+voice] or [–voice]. Thus, whichever feature is present in the laryngeal node of the second obstruent will eventually spread to the preceding one. Privative models, when confronted with symmetrical assimilation, must distinguish between a phenomenon which is a result of spreading of the active laryngeal category, in this case {L}, as in, e.g., prosić / prośba [prit ~ pr ba] ‘to ask / a request’, and one which is quite different in kind. Since Polish, under this view, utilizes only the element {L}, the assimilation to a voiceless obstruent as in dech / tchu [dx ~ txu] ‘breath, nom.sg./gen.sg.’ must be viewed as nonassimilatory in nature. Rather, it is a case of neutralization (delaryngealization) similar to word-final devoicing. The two types of assimilation are shown below.

|| 15 Gussmann (1992) differs only in the causes of delaryngealization (13a) of the first obstruent, which, for him, stem from syllabification and voice licensing.

Laryngeal Realism and Polish voicing | 15

(14)

a. Assimilation as L-spreading

Interpretation

/po r o i to/ ~ /po r o bL a/ → /LbL/ > [ b] | L [prit] [pr ba] b. Delaryngealization /dL xo/ ~ /do xo u/ |  L L [dx] [txu]

Interpretation → /doxo/ > [tx]

It appears then that surface symmetrical assimilations can be given an asymmetrical account. This has at least three very important consequences. Firstly, it becomes obvious now that the term ‘assimilation’ is a mere cover term describing the phonetic (surface) facts, which need not have a direct analogue phonologically speaking. If that is correct, then a mere observation of the presence of voice assimilation can no longer be taken as evidence for a particular phonological representation (e.g., van Rooy and Wissing 2001). Secondly, reversing the first observation a little, it is important to realize that assimilation need not involve element or feature spreading. This is particularly instructive because, as mentioned earlier in this chapter, the existence of voice assimilations is often taken as an argument that a particular system has to possess the relevant category in the first place. The third consequence concerns the predictions that such a privative analysis makes. It was noted in Lombardi (1995a: 54) that the privative analysis of voicing predicts that in the absence of delaryngealization – in our terms this would mean no L-delinking in a system like Polish – one should expect the absence of FOD and only one type of voice assimilation in obstruent clusters. In other words, /CL/ would survive in the two contexts (4b) and (4c), allowing for word-final voiced obstruents as well as clusters of obstruents which would disagree in voicing. To be precise, /CLCo/ would be found, but /CoCL/ would have to become uniform /CLCL/ because L-spreading is a different rule from L-delinking. This is exactly what seems to be found in Ukrainian in which voiced obstruents are found word-finally, e.g., [hryb] ‘mushroom’ vs. [hryp] ‘grippe’, while clusters of obstruents assimilate only to a voiced one, e.g., [proz’ba] ‘request’ (compare with the parallel Polish prośba [pr ba] ‘request’), and not to a voiceless one, e.g., Ukrainian [ridko] ‘rarely’ vs. Polish rzadko [atko] ‘rarely’. This correlation between FOD and RVA to a voiced obstruent is not

16 | Preliminaries

impossible to express in binary feature systems. It is enough to eliminate the rule of delaryngealization. However, that move would also affect [–voice] obstruents and render them immune to [+voice] spreading from the following obstruent. This is a wrong prediction. Thus, the privative systems fare better in dealing with cases like Ukrainian. Of course, rule-based approaches could still rescue the analysis of Ukrainian by specifying the rule of delaryngealization as delinking only [–voice]. This however, does not add to the explanatory value of such analyses. Returning now to Polish, it seems to allow for two configurations of obstruent clusters in phonological representation (15). (15)

a.

b. Co

Co

CL CL

They may be lexical or ‘derived’ by element spreading (/CoCL/ → /CLCL/) or element delinking (/CLCo/ → /CoCo/).16 Gussmann (2007: 291) captures this regularity by means of a Voice Adjustment principle. (16)

Voice Adjustment The tonal specification of the last obstruent controls the laryngeal tier of the sequence.

This principle is meant to account not only for cases of dynamic voice assimilation – which can be observed as processes – but also for the static phonotactic patterns found in, e.g., kto ‘who’, gdy ‘when’, etc. Later in this book, we will propose a slightly different formulation, which is based on contextual tone licensing. The conclusion we can draw from the above discussion is that the basic Polish voicing facts can be easily handled by Laryngeal Realism, allowing for interesting predictions which follow from privativity and non-specification. The three aspects of Polish voicing discussed above, that is, the distribution of contrast, final devoicing and voice agreement in obstruent clusters are common to the two major dialect groups in Polish. There is, however, one phenomenon that sets the two main dialects of Polish apart. Unfortunately, this phenomenon is inexplicable under the Laryngeal Realism view in which Polish is an L-system. One of the main problems is that sonorants appear to need a voicing property to spread in one of the dialect groups of Polish. || 16 It will be recalled that the obstruents may be separated by a sonorant. Thus, in fact, we are talking about /Co(S)Co/ and /CL(S)CL/, respectively.

The problem of Cracow-Poznań sandhi voicing | 17

6 The problem of Cracow-Poznań sandhi voicing As mentioned earlier in this chapter, Polish exhibits voicing agreement also in obstruent clusters which result across word boundaries. The agreement involves, what appears to be voice spreading in, for example, kwiat begonii [kfjad bgji] ‘begonia flower’, in which the final obstruent in the first word is lexically voiceless, cf., kwiaty [kfjat] ‘flower, pl.’, as well as in the ambiguous cases in which a lexically voiced obstruent, e.g., in szyb kopalni [ p kpali] ‘mine shaft’ is followed by a voiceless obstruent. From the point of view of Laryngeal Realism, the analysis of kwiat begonii would be parallel to that of prośba in (14a) above, while szyb kopalni corresponds to tchu (14b). This is illustrated below in (17a) and (17b) respectively, where only the relevant consonants carry the specification symbols. (17) a. Assimilation as L-spreading /k fj a to/ # /bL g  j i/ | L [kfjad bgji] b. Delaryngealization /  bo/ # /ko p a l  i]  L [ p kpali]

Interpretation → /tLbL/ > [db]

Interpretation → /boko/ > [pk]

Since the voiceless /ko/ in kopalni is unmarked, the analysis in (17b) does not use voicelessness spreading. All that needs to be said is that the final obstruent in the first word is delaryngealized. The voicing agreement between two obstruents across a word boundary is subject to some variation due to typical causes observable in this context, such as tempo of speech. Nevertheless, the facts seem to be uniform in all dialects of Polish in this respect, as can be seen in (18c,d) below. An interesting divide arises between two major dialects, that is, so-called Warsaw Polish (WP) and Cracow-Poznań Polish (CP) when the word-final obstruent of the first word is

18 | Preliminaries followed by a vowel or a sonorant consonant in the second word.17 Irrespective of the lexical marking of the final obstruent, it becomes voiced in CP in this context, and remains voiceless in WP (18a,b).18 The first observation that needs to be made is that regardless of the striking difference in terms of sandhi voicing, both dialect groups ignore the lexical distinction of its word-final obstruents. The obstruents are uniformly affected or not affected whether they were underlyingly voiced or voiceless. Thus, both groups seem to carry out some kind of neutralization in this context. This observation is important as it eliminates the possibility that, for example, the wordinitial vowel in oni or odrębny reverses the delaryngealization of the preceding word-final obstruent. This could be an option only if the sandhi voicing concerned the lexically voiced obstruents alone. Rather, we are dealing with neutralization of the laryngeal contrast in word-final position and something which looks like voicing triggered by the word-initial sonorant. Note that this happens before vowels as well as sonorant consonants. (18) a. jak oni ‘how they’ wkład odrębny ‘separate contribution’ b. jak możesz ‘how can you’ wkład mój ‘my contribution’ c. jak dobrze ‘how well’ wkład własny ‘own contribution’ d. jak trudno ‘how hard’ wkład stały ‘permanent contribution’

WP

CP

k- t- k-m t-m g-d d-v k-t t-s

g- d- g-m d-m g-d d-v k-t t-s

__V+v __S+v __C+v __C–v

Secondly, the two groups can be described as predominantly devoicing (WP) and predominantly voicing (CP). Note that in CP, the final obstruents are voiceless only in front of voiceless obstruents in the following word.19 Likewise, in WP, the final obstruents are voiced only in front of voiced obstruents and remain voiceless elsewhere.

|| 17 These major dialects are in fact dialect groups. CP includes Lesser Poland, Silesia, Greater Poland as well as Kashubia. WP involves Mazovia and the North Eastern dialects (e.g., Urbańczyk 1984). For earlier discussions of the phenomenon see, e.g., Baudouin de Courtenay (1894), Benni (1907), Nitsch (1912, 1957), Śmiech (1961). 18 Reports of similar effects concern Breton (Ternes 1970), West Flemish (De Schutter and Taeldeman (1986), Catalan (Wheeler 1986), and varieties of German and Italian (Krämer 2001). 19 Of course, they are also voiceless in absolute final position, when no other word follows.

The problem of Cracow-Poznań sandhi voicing | 19

The facts observed in WP seem to fit the Laryngeal Realism analysis rather neatly. Recall that Polish is an L-system in which voiced obstruents contain {L}, while voiceless obstruents as well as vowels and sonorant consonants are nonspecified for voice. We predict that phonological assimilation, that is, spreading of the laryngeal element, is possible only from voiced obstruents, which is exactly what we see in (18) above. In other words, the representation proposed by Laryngeal Realism provides a clear answer as to why WP has no sandhi voicing before sonorants, and why the voicing is possible before voiced obstruents. They stand out from all the other phonetically voiced segments because they are the only objects which have an active laryngeal category in their representation. To conclude, the WP facts can be accounted for in the following way. The instances of sandhi voicing, that is, both jak dobrze [jag db] < /jak # db/ ‘how well’ and wkład własny [fkwad vwasn] < /fkwad # vwasn/ ‘own contribution’ may be viewed as cases of L-spreading from the word-initial obstruent. In the case of wkład własny in (19b) we may assume that the lexically voiced /dL/ was first delaryngealized in the word-final context only to be revoiced by spreading. (19)

Sandhi voicing in Warsaw Polish a. /j a ko/ # /d b / | L

→ /kLdL/ > [gd]

b. /f k w a dL/ # ↓ /f k w a do/ # /v w a s n / | L

→ /dLvL/ > [dv]

The only way to voice an obstruent in WP is to provide it with {L}. Since only voiced obstruents possess this element in their representation, no other context may produce the voicing effect. In other words, sandhi voicing before sonorants is simply impossible in the L-system of WP. Thus, it appears that the L-system perfectly accounts for the Warsaw Polish facts. However, it fails miserably when it comes to the Cracow-Poznań data in (18a,b). It should be stressed that we are dealing here with dialects in which word-internally the voicing facts are identical – both dialects have a phonetic contrast between fully voiced vs. voiceless unaspirated obstruents, and identical voicing phenomena such as final devoicing and assimilation of voice in

20 | Preliminaries

obstruent clusters. All the above criteria are normally used to postulate an Lsystem. The data in (18) show that in CP the sandhi voicing occurs in all phonetically voiced contexts, that is, before a vowel, a sonorant consonant and a voiced obstruent. Thus, it appears that for this dialect, the voicing of obstruents is no different from the voicing of sonorants. These facts are problematic for phonological theory, regardless of whether a privative or a binary feature system is used. For example, in some privative models (e.g., Bethin 1984, 1992) the feature [+voice] must be assumed to be present in the representation of vowels and sonorant consonants at least at the relevant, late (post-lexical) level of representation in order to be manipulated by late rules. In binary feature systems (e.g., Gussmann 1992; Rubach 1996) the assimilation rules also must be kept at bay at earlier levels of derivation in order not to produce wrong results word-internally. The specification of sonorants may only become active late in the derivation and only in Cracow Polish. Most of these analytical problems with CP voicing partly stem from the assumption that it must be dealt with in terms of phonological computation, manipulating active categories such as the feature [+voice]. This need not be correct. It will be recalled that in this section we have noted that the surface presence of assimilation need not be a result of phonological spreading, as shown in (17b). Laryngeal Realism with its privative elements does not fare any better with respect to CP sandhi voicing (e.g., Gussmann 2007). If voicing in obstruents can only be due to the presence of the element {L}, then one must postulate that this element is present somewhere in the representation of the initial voiced segments in CP. This, however, goes strongly against the Element Theory in which sonorants are not specified laryngeally.20 For this reason, both Gussmann (2007) and Michalski (2008, 2009) fail to provide a viable analysis of CP voicing. In the following chapter, a modification of Laryngeal Realism is going to be proposed which takes into account the nature of the relationship between phonology and phonetics. This will be followed by a new analysis of CP sandhi voicing.

|| 20 For a proposal that sonorants have a special feature [sonorant voice] see Rice (1993).

Chapter 2 Sound system, phonology and phonetics 1 Introduction The aim of this chapter is to introduce a particular view on how phonology and phonetics interact in sound systems. The main idea is that phonetics provides an arrangement of probable phonetic contrasts, which will be called phonetic categories, for a particular number of phonological contrasts. The nature of this interaction consists in building quite arbitrary relations between categorical distinctions provided by phonology proper and their direct expression within a particular phonetic space.21 The phonetic nature of these spaces, however, is such that the interpretational phonetic choices are not entirely arbitrary, especially since in language acquisition learners do not construct the systems from scratch. They merely work out the nature of the relationship between the phonetic categories recurring in the input data with the phonological representations of the categorical distinctions to do with particular phonetic spaces.22 The phonological representations are also constructed in accordance with certain principles, for example, privativity. It will be argued that most of the confusion in the discussion on phonology and phonetics stems from the fact that sound systems are mistaken for either phonology or phonetics. First, we look at the phonetic and interpretational aspects of voicing in different types of segments and systems. This discussion follows and further develops the proposals in Harris (2009) concerning what he calls ‘the disunity of voice’. There, the main argument is that voicing depends on three types of aspect: i) segment types, ii) phonological contexts, and iii) languages. Thus, concerning the first aspect, voicing is different in kind between sonorants and obstruents, which, as is often claimed especially in privative studies, should be reflected in a different phonological treatment. While we will generally agree

|| 21 Although our direct interest is in the nature of laryngeal systems, other phonetic dimensions can constitute domains of such interaction, such as place (including the palatalization complex), manner, and the vowel system. 22 To see how particular phonetic spaces correspond to ‘contrastive hierarchies’ in phonological representation see, e.g., Dresher (2003).

22 | Sound system, phonology and phonetics

also with the remaining two aspects, we will emphasize the distinction between phonological and phonetic contexts and propose an extension to the division into ‘spontaneous’, ‘active’ and ‘passive’ voicing, by proposing a fourth, interpretationally based type of voicing, which seems to contradict Harris’s carriersignal modulation model. This chapter introduces the structure of sound systems, with a particular focus on laryngeal systems, using the Government Phonology view of what the phonological side of the equation is. This is followed by a discussion of phonetic interpretation and a proposal to replace Laryngeal Realism with Laryngeal Relativism, which seems to be necessary if the problem of Cracow-Poznań sandhi voicing signalled in the previous chapter is to receive a satisfactory analysis within a privative model. The chapter ends with a discussion of some consequences of Laryngeal Relativism in anticipation of the application of this model to the basic Polish voicing data as well as the intriguing phenomenon of CP sandhi voicing, for which a new analysis will be provided in the following chapter.

2 Types of voicing Depending on the different types of segments, that is, whether we are dealing with a sonorant consonant or an obstruent, as well as on the different phonological status of the obstruents within a particular laryngeal system, that is, marked or non-specified, we may distinguish three different situations to do with voicing. A fourth one will be added after Laryngeal Relativism has been introduced.

2.1 Spontaneous voicing If a sufficient drop in air pressure and air flow between the trachea and pharynx (across the glottis) is maintained vocal cords are subject to what is called spontaneous vibration (voicing). Next to these aerodynamic conditions, some other articulatory parameters must be met. For example, the vocal cords should not be spread apart and not tensed (e.g., Chomsky and Halle 1968; Halle and Stevens 1971). This situation typically occurs in unoccluded speech sounds such as vowels and sonorant consonants which, for reasons of maintaining an open articulation, are not characterized by an intra-oral air pressure build-up. However, obstruents can also be spontaneously voiced under certain conditions. These involve the same aerodynamic characteristics, as well as a number

Types of voicing | 23

of articulatory parameters in addition to those mentioned above, which allow for the state in which vocal folds vibrate spontaneously in stops (e.g., Westbury and Keating 1986: 151). Some of these parameters involve, for example, relatively short closure, contracting the respiratory muscles, decreasing the average area of the glottis and / or tension of the vocal folds, decreasing the level of activity in muscles which underlie the walls of the supraglottal cavity, active enlargement of the volume of that cavity, etc. In other words, the vocal fold vibration can be orchestrated by a number of means, and there seems to be no one direct gesture causing the vibration.23 Unlike in sonorants, an important aspect of spontaneous voicing in obstruents is connected with the adjacent phonetic context.24 Some contexts are more conducive to spontaneous voicing maintenance than others (e.g., Westbury and Keating 1986; Harris 2009). For example, an inhibitory effect on such natural voicing is typically observed in word-initial and word-final contexts, as opposed to the intervocalic context, which allows for a maintenance of the articulatory state connected with vocal fold vibration. Westbury and Keating (1986: 163) note that, quite surprisingly, it is more ‘natural’ in terms of articulation to have a voiced stop in intervocalic contexts than a voiceless unaspirated one, as the latter requires a change in a steady articulatory state – switching off the vibration of the vocal folds. However the voiceless unaspirated realization is favoured in pre-contrast systems, that is, systems with no phonological contrast based on voice, despite the fact that it requires a greater articulatory effort. Westbury and Keating (1986) explain this interesting paradox, by referring to what they call ‘more powerful principles’ governing phonetic interpretation in sound systems whereby a voiceless unaspirated stop is maintained in intervocalic position, even though it would be phonetically more natural, in the sense of ease of articulation, to spontaneously voice the stop in that position. They give the example of English, a two-way laryngeal system, in which speakers may produce voicing in word-initial /b,d,g/ even though it is not necessary for the maintenance of contrast because the other type of stops, /p,t,k/, are articulated with aspiration in that position. They conclude that next to phonetic naturalness there is a systemic tendency to maintain the phonetic similarity among the positional allophones. || 23 If a number of articulatory gestures can be used to achieve the state of vocal fold vibration, and the voicing contrast is usually manifested by a number of acoustic / auditory cues, it is clear that representing the contrast by just one phonological category must lead to a fairly arbitrary relationship between phonological and phonetic categories. 24 This does not mean that voicing might not be inhibited in sonorants in certain voiceless environments. This issue will be developed further in Chapter 5.

24 | Sound system, phonology and phonetics

It seems that in the case of English we are dealing not only, or not so much, with the maintenance of phonetic similarity among the positional allophones, as with an interpretational decision encompassing both stops and fricatives. The word-initial voicing appears to be more important in the case of the fricatives where aspiration is less robust a cue, and the voicing allows for the preservation of sufficient discriminability between the two laryngeally contrasting series. We may conclude the discussion of spontaneous voicing by stressing a number of points. Natural phonetic interpretation of speech sounds in terms of voicing may be quite directly connected with the general aerodynamic and articulatory settings. However, this would apply only to the spontaneous voicing in vowels. The situation becomes a little more complicated with sonorant consonants, if they find themselves in voiceless environments. For example, in Polish, sonorant consonants tend to be devoiced if surrounded by voiceless obstruents, e.g., krtań [krta] ‘larynx, nom.sg.’, or in word-final position, when preceded by a voiceless sound, e.g., wiatr [vjatr] ‘wind, nom.sg.’. However, in the latter cases, we may still talk about ‘natural’ phonetic interpretation involving contextual inhibition of spontaneous voicing. With obstruents, on the other hand, the situation is much more complicated. Firstly, the retention of a voiceless unaspirated stop in the context in which it would be pronounced more naturally as voiced suggests that the ease of articulation cannot be viewed as the only factor in sound patterns. Secondly, and more importantly, for our purposes, there seem to be ‘more powerful principles’ or systemic decisions as to the phonetic shape of one of the obstruents in a system possessing a two-way contrast, but also in pre-contrast systems. One thing that we can say about these principles at this point is that they are not purely phonetic. Rather, they are systemic decisions on phonetic interpretation of segments, which are clearly dependent on the type of the phonological representation with respect to voicing. It follows from the above discussion that spontaneous voicing may occur not only in sonorants but also in obstruents under certain special circumstances. In a sense, given the indirectness of the aerodynamic conditions and the articulatory parameters in connection with vocal fold vibration, it is possible to claim that all voicing is in fact spontaneous. However, for the sake of the following discussion, we will restrict the use of the term spontaneous voicing to vowels and sonorant consonants. For obstruents, we will distinguish two more types of voicing, ‘active’ and ‘passive’, which are meant to express the different situations involved. In the following sections, we will introduce another type of voicing which will follow from the proposed model of Laryngeal Relativism.

Types of voicing | 25

2.2 Active voicing We follow, e.g., Harris (2009) in assuming that active voicing in obstruents is connected with the interpretation of the laryngeal element responsible for voicing, that is, {L}. To use an argument to do with the aerodynamics of voicing, active voicing is a kind of compensation for the intra-oral air pressure build-up arising due to obstruent stricture, which has an inhibitory effect on vocal fold vibration, by means of an active gesture or a set of gestures offsetting this effect. Viewed from the perspective of the relation between the phonetic fact of vocal fold vibration and the representation of laryngeal distinctions, in such cases there seems to be a one-to-one correspondence between the presence of the voicing cue in the signal (negative VOT) and the presence of the phonological category {L} in the representation. In the modulated-carrier model of speech, to which Harris (2009) adheres, features map onto modulations of the carrier signal. In this case, for example, the presence of {L} entails the full voicing (long VOT lead) in initial plosives. In this respect, his model is essentially no different from the Laryngeal Realism approach. Like Laryngeal Realism, the modulated-carrier model also recognizes the existence of ‘passive’ voicing, to be discussed next. This suggests that not every case of carrier modulation is related to the presence of an active phonological category. The listener needs to be able to distinguish between periodicity that is part of the carrier signal and periodicity that represents a linguistically significant modulation. (Harris 2009: 32)

It will be recalled from the previous chapter, that in Laryngeal Realism (Harris 1994; Honeybone 2002), laryngeal specification in two-way contrast systems is privative. One of the series of obstruents is marked, while the other remains non-specified. In systems with active voicing it is the voiced series that contains {L}, and the non-specified obstruents are voiceless unaspirated. Thus, Polish is an example of a system with active voicing. It is interesting to note at this point that in such a system, the systemic decisions (‘more powerful principles’) concerning the interpretation of the non-specified congener in fact forbid its spontaneous voicing. This point will be further developed in the following sections.

26 | Sound system, phonology and phonetics

2.3 Passive voicing According to, e.g., Kohler (1984: 162), passive voicing occurs in the unmarked obstruents in which the aerodynamic conditions for voicing are said not to be actively controlled. In the previous section, we saw that this type of voicing is forbidden in L-systems. This means that the aerodynamic conditions of the neutral series in an L-system must be actively controlled to prevent passive voicing. Passive voicing is restricted to systems like English in which the so called voiced obstruents are in fact phonologically unspecified and the voiceless aspirated stops are marked with a laryngeal category (Avery and Idsardi 2001; Harris 1994, 2009; Iverson and Salmons 1995, 2003a; Kohler 1984). In terms of the Element Theory and the Laryngeal Realism approach, this passive voicing occurs in H-systems, that is, in ‘aspiration’ languages. An additional condition on passive voicing, it will be recalled, is the presence of voiced environment. A rather restrictive definition of passive voicing in terms of context is given in Iverson and Salmons (2003a: 51). The authors view it as an extension of spontaneous voicing from a preceding segment onto an unmarked obstruent. Thus, for example, the rightward influence allows Iverson and Salmons to view the /b/ in rubber and the /d/ in bad as passively (post)-voiced. However, the /b/ in bad cannot be viewed as a case of passive voicing in those varieties of English in which it is voiced. In the light of our discussion in the previous section, Iverson and Salmon’s decision is intuitively correct for English. Word-initially, the voicing of stops goes beyond what seems to be necessary, and some varieties of English do not even show it. It should be emphasized, that unlike in sonorants, in which spontaneous voicing is also associated with the absence of active control of the aerodynamic conditions, passive voicing in obstruents is not of the same kind. Firstly, there are systemic conditions that must be fulfilled. It is required that the obstruent is non-specified for voice, and it is part of a system with a two-way laryngeal contrast in which the marked segments carry {H}, that is, the category responsible for aspiration. Secondly, crucial in this type of voicing, is a voiced environment as well as an articulation which is conducive to maintaining the air pressure drop across the vocal folds. We may identify this articulation as weak, or lenis. When we consider these conditions, it turns out that passive voicing is not just phonetic in nature. It is, in equal measure, an interpretational phenomenon which is dependent on the particular phonological marking in a given system. In other words, it is a systemic decision (e.g., contrast enhancement) that leads to lenis articulation. This intuition seems to be correctly expressed not only in

Types of voicing | 27

Westbury and Keating’s (1986) ‘more powerful principles’ which were mentioned above, but also in Iverson and Salmons’s (2003a) claim that passive voicing is in fact a shallow phonological rather than purely phonetic fact. The fine distinction between phonetics and phonetic interpretation will be developed further below. The articulatory parameters in lenis obstruents of some ‘aspiration’ languages are in a sense actively set to be weak. This point becomes particularly clear when we consider the interpretation of the unmarked series in a voicing language, that is, in one that marks full voicing with a category, e.g., {L}. Here, passive voicing is simply forbidden, as it would be difficult to distinguish between fully voiced and weakly voiced obstruents, let alone produce such a contrast. In this case, the contrast enhancement is less paradoxical as it does not lead to a lenis articulation of the non-marked obstruents, but to one that will guarantee the absence of passive voicing. To conclude, passive voicing is a term describing a situation in which we are dealing with phonetic and interpretational voicing of the unmarked obstruents in an H-system (aspiration language). From the interpretational point of view, passive voicing may involve an active decision, a kind of enhancement of the categorical distinction if one prefers. It is, indeed, a paradoxical situation, when enhancement is implemented by weakening of the articulation. However, once we assume that weakening is one of the means of achieving contrast enhancement, the contradiction disappears. Two other conclusions follow from the above discussion. There is a need to clarify the nature of the systemic decisions mentioned above and the structure of the system itself with respect to the interaction between phonological representation and the phonetic facts. One interesting finding is that we should make a distinction between what might be called universal phonetic principles, like spontaneous voicing which relates to the physiology of speech and to phenomena which can be studied independently of phonology, and principles of phonetic interpretation, whether universal or systemic, which take into account both the phonological and the phonetic aspects. With respect to voicing, the difference becomes clear when we consider the types of spontaneous, passive, and active voicing in the table below. We include the information which follows from the three aspects of the ‘disunity of voice’ that Harris (2009) emphasizes, that is, i) the type of segment involved, ii) the phonological and phonetic context, and iii) the type of system in which particular segments occur. It appears that while talking about the ‘disunity of voice’ we begin to observe a kind of continuum, of which we are going to take full advantage in the following chapters, in which we discuss two puz-

28 | Sound system, phonology and phonetics

zling phenomena concerning the Polish voicing system, namely, pre-sonorant sandhi voicing and the so called progressive devoicing. (1)

Voicing continuum voiced universal phonetic principles spontaneous voicing Vo

So

vowels

sonorant consonants

unconditioned

contextually conditioned

universal and systemic principles of phonetic interpretation passive voicing

active voicing

Co

CL obstruents

systemically and contextually conditioned

phonologically conditioned

‘aspiration’ systems (Co vs. CH)

‘voicing’ systems (Co vs. CL)

The types of segments in the above table seem to constitute a scale from the laryngeally unmarked vowels which are least conditioned cross-linguistically in terms of voicing, through slightly more conditioned sonorant consonants to highly conditioned obstruents, of which some are still unmarked phonologically, while others carry the element {L}. It is clear that universal phonetic principles end with sonorants and that obstruents are subject to the ‘more powerful’ principles of phonetic interpretation.25 In the following sections, a fourth type of voicing situation will be proposed. We proceed now to the discussion of the structure of sound systems.

|| 25 These principles, for example, ensure that the phonologically identical object /Co/ will function differently depending on the type of system it finds itself in. For example, in an Lsystem, /Co/ must not be allowed to be passively voiced, while in an L-system, /Co/ may but does not have to be voiced passively.

Sound system | 29

3 Sound system The concept of ‘sound system’ is crucial in this work. A system, then, including a laryngeal system, may be defined as the sum of phonological, interpretational, and phonetic aspects which together are responsible for the observed phonetic facts, which we will refer to as sound patterns. Let us look at a schematic illustration of the structure of sound systems as understood in this work. (2)

Sound system (e.g., Laryngeal system) Phonology

Phonetics

Representation & Computation

Phonetic categories & Phonetic interpretation

- privative categories - (un)licensing, government - (de)composition: spreading, delinking

- universal principles - system specific conventions - sociolinguistic modifications

It is a view, in which phonology and phonetics are kept strictly apart, yet they form two sides of the same coin and are mutually dependent, to the extent that it is quite impossible to study sound systems without making reference to both phonology and phonetics as well as to the way these domains interact within a particular system. As can be seen in the diagram in (2), a sound system is the sum total of phonological and phonetic aspects. Phonetic facts, or sound patterns, follow from the interaction between the two sides of the coin. More specifically, it will be argued, there is a direct interaction between phonological representation and phonetic interpretation principles, rather than, between representation and phonetic categories. The latter option is represented in the Laryngeal Realism tradition, which has been shown earlier to work only for one dialect group in Polish. Given the above structure, a sound system cannot be identified only with phonology, or only with phonetics. Sound patterns are always a result of system dependent phonetic interpretation of phonological representation. They follow from the system, and as such they may be ambiguous and misleading. In other

30 | Sound system, phonology and phonetics

words, to understand a sound pattern one has to work out the sound system behind it, that is, how phonology and phonetics interact in that system. Both phonology and phonetics are separate and can be studied separately, but when sound patterns or systems are taken into account, the two aspects must dove-tail to produce the results. Below, the phonological and phonetic sides of the system are discussed in more detail.

3.1 Phonological aspects In short, phonology involves abstract symbols (representation) and principles of their distribution and manipulation (computation). Since our primary interest lies in the laryngeal system of Polish, that is, the melodic rather than formal aspect of representation, we restrict the discussion to elements responsible for defining laryngeal contrasts, that is, {L} and {H}, and follow Laryngeal Realism until the modifications are fully introduced. One of the functions of phonology is to define categorical contrast. In privative models this boils down to a presence or absence of a particular property to distinguish segments from one another. The privativity of phonological categories has been extensively argued for in the literature (see, e.g., Avery 1996; Harris 1990, 1994, 2009; Honeybone 2002, 2005; Iverson and Salmons 1995; Lombardi 1991, 1995a). We assume a standard set of elements known from the Government Phonology (GP) literature, that is, {I}, {U}, {A}, {H}, {L}, {h}, { }, {N}, adapted from, e.g., Harris (1994). Each element is independently pronounceable and so is any combination of elements. For example, in vowel systems, the corner vowels /i, a, u/ correspond to the resonance elements {I}, {A}, {U}, while mid vowels /e, o/ are compound structures {I-A} and {U-A} respectively. The resonance elements define place in consonants: {I}-palatality, {U}-labiality, {A}coronality, or just simplex consonants, {I}=/j/, {U}=/w/, {A}=/r/.26 As for computation, GP recognizes two operations: composition and decomposition. For example, mid vowels may be decomposed in unstressed position, as in Bulgarian {I-A} → {I}, corresponding to [e] → [i]. The opposite, that is, composition, is a reverse process in which an element is fused with another one under some conditions. One such condition is the local presence of the added element. Such phenomena are observed under the rubric of, e.g., vowel coalescence or vowel harmony.

|| 26 The fact that more recently some of these elements have been abandoned, conflated, or replaced by other, e.g., structural means, does not have any bearing our discussion.

Sound system | 31

Government and licensing organize the phonological representation through lateral relations that segments contract with each other (Cyran 2010; Kaye, Lowenstamm and Vergnaud 1990; Scheer 2004).27 One of the main principles determining the distribution of melodies in phonological representation is licensing. Thus, for example, segment decomposition typically occurs under weak licensing, which correlates with weak prosodic contexts. For vowels, it is the unstressed position which has weak licensing. On the other hand, for consonants, such context has been identified as the coda position.28 The processes of Final Obstruent Devoicing (FOD) and Regressive Voice Assimilation (RVA) discussed in the previous chapter are good examples of decomposition and composition with respect to voicing phenomena. These are repeated below, with licensing added to the picture. The precise cause of weak licensing in this position will be addressed in detail in Chapter 5. (3)

a. RVA

b. FOD weak licensing

l



t |

o

+

/bowtm/

[bwtm] błotem ‘mud, instr.’

WP CP

VCoV VCHV

/rsoa/ /rsHa/

>

[rsa]

rysa ‘scratch’

WP CP

VCLV VCoV

/rzLa/ /rzoa/

>

[rza]

ryza ‘ream’

WP CP

VCoSV VCHSV

/ko/ /kH/

>

[k]

oknie ‘window, loc.’

WP CP

VCLSV VCoSV

/gL/ /go/

>

[g]

ognie ‘fire, pl.’

The fully voiced obstruents, on the other hand, are non-specified in CP and marked with {L} in WP (e.g., WP /rzLa/ vs. CP /rzoa/ > [rza]). These representation-to-interpretation relations are schematized below in a different type of graph. (3)

WP

[b] CL

CP

[p] Co

[b]

CH

[p]

As shown in (3), the phonetic categories in the two dialect groups are the same, respecting the same phonetic distance. However, these phonetic categories belong to two different systems: with opposite phonological representations, and consequently, also disparate phonetic interpretation conventions. Let us now turn to the description of the behaviour of the two systems in neutralizing

Laryngeal Relativism and Polish voicing | 47

contexts, that is, /__#/ and /__C/, where we observe FOD and RVA. It will be shown that the two systems are able to generate the same phonetic facts.

2.1 Final Obstruent Devoicing It will be shown presently that despite the reversed representation assumed for CP and WP, not only the surface facts are identical, but also the phonological computation formally involves the same process, that is, delaryngealization. It will be recalled that computationally, FOD in Warsaw Polish involves Ldelinking.39 Thus, for CP, we propose that in the same context, the element {H} is also delinked (4b). (4)

a. L-delinking in WP CL → Co / __#

b. H-delinking in CP CH → Co / __#

The systemic interpretation of the resulting /Co/ in WP is that of a voiceless unaspirated obstruent. As for CP, we saw above in (2) and (3) that in nonneutralizing contexts the interpretation of /CH/ is a voiceless unaspirated obstruent, e.g., (/pH/ > [p]), while /Co/ is realized as a fully voiced obstruent, e.g., (/po/ > [b]). The question is what to expect from CP /Co/ in the neutralizing context, and whether the delaryngealization is at all necessary in that dialect group, since /CH/ is responsible for the voiceless unaspirated rendition – exactly what we get in that context. First of all, let us recall that ‘passive’ voicing in obstruents requires a voiced phonetic context. We assume that the ‘enhanced passive’ voicing in CP is subject to the same conditions. Specifically, a spontaneously or actively voiced segment is needed in the following context. We observed this in the nonneutralizing context /...C(S)V.../ in (2). Consequently, /Co/ in the word-final context cannot be voiced, which produces a result of voicelessness. Here, we are simply dealing with the absence of phonetic voicing, and not with FOD proper. Our reasons for postulating H-delinking in CP in the word-final context will become apparent presently. It should be noted however, that one should expect

|| 39 Details concerning the causality of this kind of delaryngealization, as well as of the one in front of obstruents (__C), which leads to voice assimilation in obstruent clusters, will be discussed in the first section of Chapter 5.

48 | Cracow-Poznań sandhi voicing

a phonetic distinction between the non-voiced /Co/ and the marked /CH/ in that context, if there was no delaryngealization.40 Let us compare the derivations and interpretations below, in which WP and CP are compared with respect to the behaviour of their final obstruents. First, we look at the treatment of the so called Final Obstruent Devoicing in (5) in the alternating pair żaba / żab [aba ~ ap] ‘frog, nom.sg./ gen.pl.’. We assume that there may be a difference between the so called lexical and phonological representation. The difference is due to the operation of phonological processes, such as spreading and delinking, which are caused by different licensing conditions resulting from morphological activity, e.g., suffixation. If no phonological process takes place, then the phonological representation is identical to the lexical one. Thus, the lexical representations are also phonological representations, which can be interpreted phonetically just like the latter. The following symbols are used: ‘→’ a phonological process, ‘=’ no change between lexical and phonological representations, ‘>’ phonetic interpretation. lexical representation

(5)

phonological representation

phonetic interpretation

a. Warsaw Polish /LabL-/

FOD

/LabL-/

=

/LabLa /

L-delinking → /Labo/

>

[aba] ‘frog, nom.sg.’

>

[ap]‘frog, gen.pl.’

b. Cracow Polish /oabo-/

=

/oaboa/

>

[aba] ‘frog, nom.sg.’

‘FOD’ /oabo-/

=

/oabo/

>

[ap]‘frog, gen.pl.’

Both in WP and in CP, all the interpretational distinctions are maintained in the nominative, because it is a non-neutralizing context – before a vowel. Thus, [b] is the interpretation of /bL/ in WP (5a), and of /bo/ in CP (5b). It should be

|| 40 The problem of incomplete neutralization in the word-final context will be returned to in the final chapter. Brockhaus (1995) in fact proposed that in German we are dealing with no delarygealization of /CH/ and phonetic non-voicing of the delaryngealized /Co/. The latter, however, comes from /CL/.

Laryngeal Relativism and Polish voicing | 49

stressed that the voiced symbol /b/ is used here to express the fact that these objects will be fully voiced in the non-neutralizing context. The FOD context brings out the crucial difference between WP and CP. In the former, final devoicing is a case of delaryngealization whereby the element {L} is delinked domain-finally (see, e.g., Gussmann 2007), and the neutral (or neutralized) obstruent in the L-system is phonetically interpreted as voiceless unaspirated. Thus, in WP, we are dealing with a truly phonological FOD – the phonological representation is manipulated by computation. This is followed by interpretation of the resulting structure. In CP, on the other hand, which is an H-system, there is no phonological computation in the alternating pair in (5b). In the nominative, /Co/ is phonetically realized as a fully voiced object. On the other hand, the word-final context disallows passive voicing and the result is ‘FOD’. The inverted commas express the fact that in Cracow Polish there is no phonological FOD, there is phonetically based non-voicing of the neutral object (e.g., Brockhaus 1995; Harris 2009). This is due to the difference in the way /Co/ is systemically interpreted in the two contexts shown below. (6)

Phonetic interpretation of /Co/ in CP a. /Co/ > [voiced]

/ __V

b. /Co/ > [voiceless] / __# The rule format used above is merely a handy convention to schematize the systemic interpretation conventions in a given dialect. Note that the conventions above use the symbol ‘>’, which means: ‘phonetically interpreted as’, rather than ‘→’ which means: ‘phonological representation A is changed into phonological representation B’. It should be added that the two conventions above have both phonetic and systemic grounding in CP. Let us now compare WP and CP with respect to the voiceless unaspirated obstruents in the neutralizing and the non-neutralizing contexts. The pair of words is mapa / map [mapa ~ map] ‘map, nom.sg./ gen.pl.’. The situation concerning the voiceless unaspirated obstruent in WP is rather simple. [p] is always a phonetic interpretation of /Co/. In CP, on the other hand, [p] is a phonetic interpretation of /CH/ in the non-neutralizing contexts, and of /Co/ word-finally. A lexical /CH/ is assumed here not to survive in the word-final context, just as {L} could not survive in this context in WP.

50 | Cracow-Poznań sandhi voicing

(7)

lexical representation

phonological representation

phonetic interpretation

a. Warsaw Polish /mapo-/

=

/mapoa/

> [mapa] ‘mapa, nom.sg.’

/mapo-/

=

/mapo/

> [map] ‘map, gen.pl.’

=

/mapHa/

> [mapa] ‘mapa, nom.sg.’

b. Cracow Polish /mapH-/

/mapH-/

H-delinking → /mapo/

> [map] ‘map, gen.pl.’

Thus, both dialect groups have a phonological neutralization word-finally in that only /Co/ is allowed in this context, but we are dealing with quite disparate systems with respect to the status of FOD. Final devoicing in an H-system is in fact a misnomer. It is rather a case of absence of passive voicing in front of silence. Word-final context inhibits passive or natural voicing in obstruents (Westbury and Keating 1986; Harris 2009). Thus FOD in CP is an interpretational rather than phonological phenomenon. It is a consequence which, if correct, is another argument against the common practice of making an automatic equation between a phonetically observed fact and its phonological status. It will be recalled that in analyses of laryngeal systems, it is common practice to use the presence of phenomena such as assimilation and final devoicing as a diagnostic for the analytic decisions, e.g., concerning the representation of voice. In privative models, FOD is taken as an argument that a phonological category [voice] is present in the system and it is delinked in the relevant context. In other words, FOD tends to be viewed as having a phonological basis. In this context, it is a good moment to ponder on the fallacy of textbook analyses of voice alternations in word-final position which take any such instance as a case of phonological devoicing rather than lack of passive voicing. Given an alternation of the type żaba / żab aba ~ ap] ‘frog, nom.sg./ gen.pl.’, the typical argumentation is that we are dealing with final devoicing in [ap]. Alternatively, we would have to claim that the voiced [b] in aba] is a case of word-medial voicing. This in turn, the argument goes, cannot be correct because such intervocalic voicing would eliminate forms with voiceless obstruents in this context. In other words, instead of mapa [mapa] ‘map, nom.sg.’, we would expect to hear *[maba] with intervocalically voiced [b]. It is clear from our discussion in this chapter that the argumentation is false. Firstly, intervo-

Laryngeal Relativism and Polish voicing | 51

calic voicing can only take the form of passive voicing of /Co/ in an H-system, that is, a system in which the voiceless congeners carry {H}. Secondly, the tone element must be lost in order to allow for such passive voicing. This, however, does not happen in the intervocalic context, because it is not a neutralizing one. The following vowel licenses the presence of {H} in /mapHa/. Thus, the didactic value of such argumentation is rather doubtful, as it does not look at the type of system within which a particular phonetic effect is observed, and, in fact, it confuses phonetically observed facts with phonology proper. We cannot expect one and the same vowel to license the laryngeal element {H}, responsible for voicelessness in the preceding obstruent, and at the same time induce enhanced passive voicing in that object. The obstruent which is lexically specified with {H} may be passively voiced only if it first loses its laryngeal element. We will see one instance of this prediction in the following discussion. Note, at this point, that in an L-system, like WP, passive voicing of the unmarked objects is also impossible due to universal interpretational reasons: no passive voicing occurs in a language employing {L}. To conclude the discussion of final devoicing in the two dialects of Polish, the opposite phonological representation of the laryngeal contrast allows for obtaining the same phonetic facts. Namely, the phonetic contrast is maintained before a vowel, and lost word-finally in both dialects. However, in WP, FOD is a reflection of a true phonological process of L-delinking, while in CP, it is merely an interpretational phenomenon. Thus, the opposite representation must have different principles of phonetic interpretation. These principles of phonetic interpretation refer to system-based decisions, such that an L-system must not have passive voicing of the neutral obstruents, and an H-system with no aspiration (VOT lag) must have enhanced passive voicing of the neutral obstruents (long negative VOT). These system-based decisions are not exactly universal interpretations, such as defaults of the type: [sonorant] > [voiced]. Note that in this approach there is no universal default: [obstruent] > [voiceless], especially in two-way contrast systems, as these necessarily involve a systemic, not universal decision. Additionally, we postulate, that in CP, the element {H} in word-final position will be unlicensed and delinked, just as the element {L} is delinked in WP. Computationally, then, the phonologies of CP and WP are identical, except that a different laryngeal element is lost. This delaryngealization in an H-system has no phonetic consequences as long as the simplified consonant is pre-pausal, because a neutralized /po/ in map will obviously have the same phonetic interpretation as the lexically neutral /bo/ of żab in that dialect, that is, voiceless unaspirated due to absence of the phonetic conditions for interpreting such an object as voiced. The consequence of the delaryngealization in the H-system

52 | Cracow-Poznań sandhi voicing

will become obvious presently. But first, let us see how Regressive Voice Assimilation can be handled in the two systems.

2.2 Regressive Voice Assimilation WP and CP have more or less the same contrast distribution within the word and identical phonetic effects connected with the right edge of words, if these words are considered in isolation. We know now that two disparate systems may stand behind these uniform effects. Let us compare the two systems in how they deal with the voice agreement facts in Polish. The privative tradition in laryngeal phonology introduced a subtle distinction in the analysis of such symmetrical voice assimilation facts as the ones observed in Polish (e.g., Lombardi 1995a, 1995b). We saw in Chapter 1 that privative representation of the opposition enforces an analysis in which surface symmetrical assimilations must be treated as two different phenomena: i) true phonological assimilation due to spreading of an active category, and ii) delaryngealization. In this respect it can be expected that the voice assimilation phenomena within the phonological word in CP will look almost identical to what was shown for WP in Chapter 1. The crucial difference will lie in the reversed marking enforcing a reversed interpretation of the facts. As a result, the voicing agreement between two obstruents will be expressed by two types of structures. The first type will involve two neutral obstruents (8a), where the agreement has a lexical origin (/CoCo/), or is derived by delaryngealization of the first obstruent (/CLarCo/ → /CoCo /).41 The second type of structure, will have the same laryngeal element linked to both members of the cluster (8b). The latter structure may also be either lexical (/CLarCLar/), or derived, this time by spreading of the laryngeal element from the second member to the first one (/CoCLar/ → /CLarCLar/). (8)

a. Co

Co

b. C

C L/H

Below, we look at how the above configurations correspond to real examples, beginning with cases of static agreement, that is, obstruent clusters which are lexically uniform. The lexically non-specified obstruents will, of course, corre|| 41 /CLar/ stands for either /CL/, or / CH/, depending on the system.

Laryngeal Relativism and Polish voicing | 53

spond to different voicing facts because the phonetic interpretation of /Co/ differs between WP and CP. (9)

Voice agreement of neutral obstruents a. WP

b. CP

k t  | | | Co Co V kto ‘who’

g d  | | | Co Co V gdy ‘when’

As predicted by the systemic interpretation conventions operating in the nonneutralizing (prevocalic) context, in WP the cluster of non-specified obstruents will be voiceless (9a). On the other hand, in CP it will be voiced. We assume, then, that the enhanced passive voicing of the prevocalic [d] is a context for the same type of voicing of the first obstruent, a simple case of anticipatory assimilation, which is common and expected in a system with long negative VOT in its voiced obstruents. This interpretation of the voicing agreement between neutral obstruents enforces a different perspective on voice assimilations. It seems that systems with long negative VOT will exhibit assimilation to this voiced object regardless of whether this phonetic value is due to an active laryngeal category, that is, active voicing and element spreading, or whether it is simply a case of enhanced passive voicing, as in CP.42 The sequence /godo/ will be pronounced as voiceless in CP only when no vowel follows, e.g., szmaragdy / szmaragd [ maragd ~ marakt] ‘emerald, nom.pl./ nom.sg.’. Let us now turn to the voice agreement of marked obstruents. (10)

Voice agreement of marked obstruents

a. WP

b. CP d  | | C V

g | C L

k t  | | | C C V H

|| 42 Similar conclusions are drawn with respect to other languages which might behave like CP, namely, Dutch, Yiddish, and Afrikaans (van Rooy and Wissing 2001).

54 | Cracow-Poznań sandhi voicing

As can be observed, the same two words from (9) above are used for the illustration, except that their representations are reversed for the respective dialects in accordance with their respective laryngeal systems. As mentioned above, the configurations in which the two obstruents are lexically marked by one laryngeal category are no different from the phonological representations which are derived by spreading of the active category, that is, in dynamic assimilation, to which we refer as Regressive Voice Assimilation. This is illustrated below on the basis of the familiar examples prosić / prośba [prit ~ pr ba] ‘to ask / a request’ and dech / tchu [dx ~ txu] ‘breath, nom.sg./ gen.sg.’. The former is a case of regressive assimilation with a voiced result, while the latter produces a voiceless cluster. It will be recalled from Chapter 1 that the respective assimilations in an L-system, like Warsaw Polish, when viewed in privative terms, involve Lspreading in prośba [pr ba] ‘a request’ and pre-obstruent delaryngealization in tchu [txu] ‘breath, gen.sg.’. As shown below, the analysis of RVA in CracowPoznań Polish will be identical, except that the types of phonological computation involved in the respective cases are reversed. Thus, the phonological computation is identical, except that the representations impose their distinct application. (11)

Dynamic assimilation (RVA) WP a.

CP

spreading

delinking

p r  b a p r  43 b a | | | | | | C C V Co CL V →

p r  b a p r  b a | | | | | | CH Co V → Co Co V

L b.

delinking d x u | | | CL Co V →

spreading t x u | | | Co Co V

d x u | | | C o CH V →

t x u | | | C C V H

|| 43 For expository reasons we change the symbol to a voiced / /, to include the phonetic interpretation in these diagrams. Phonologically speaking it is // with the element {L} spread onto it, that is, in fact, an object which is identical to the lexical voiced / L/.

Laryngeal Relativism and Polish voicing | 55

The derivation of [pr ba] in the two dialects takes two different paths, as illustrated in (11a). In WP, the consonant sequence is composed of an unmarked object followed by a marked one. The laryngeal element spreads from the second obstruent to the first one, and the thus derived new phonological representation is interpreted as a voiced cluster (/...obLa/ → /...LbLa/ > [... ba]). In CP, on the other hand, it is the first obstruent in the sequence that is marked lexically, but it is in the context for delaryngealization (__C). The resulting cluster of neutral obstruents is realized phonetically as voiced (/...Hboa/ → /...oboa/ > [... ba]). Turning now to the derivation of [txu] (11b), we know that the first obstruent is lexically distinct from the following one, and it should yield a voiced segment if a vowel intervenes ([dx]). The voiced [d] translates into a marked segment /dL/ in WP and an unmarked /do/ in CP. Thus, in WP we are dealing with delinking of the laryngeal element {L} in [txu], and a voiceless interpretation of the cluster (/dLxou/ → /doxou/ > [txu]), while in CP, it is a case of H-spreading (/doxHu/ → /dHxHu/ > [txu]). The fact that the two systems work in a mirrored fashion does not prevent them from producing generally identical phonetic facts. As in the case of FOD, we observe reversed phonological representations, identical phonological computation to do with active categories, and reversed, or opposite phonetic interpretation principles. The interpretation of obstruents with respect to voicing is systemic. In Polish, the neutral obstruent (/Co/) is phonetically realized as voiceless unaspirated in the Warsaw dialect and as fully voiced in Cracow. Note, also, that the RVA facts lend some support to the assumption that both dialect groups have delaryngealization in the same contexts. Recall that this was not so obvious in the case of FOD, where one could claim that in the word-final context, CP allows for two distinct representations of obstruents, that is, /Co/ and /CH/ yielding virtually the same phonetic objects due to the contextual absence of enhanced passive voicing of the neutral representation. Such an analysis would probably be happily embraced by analysts who observe significant phonetic differences between the lexically ‘voiced’ and ‘voiceless’ obstruents (e.g., Slowiaczek and Dinnsen 1985).44 We will return to this discussion in the final chapter. The RVA facts, on the other hand, suggest that delaryngealization works in the same fashion in both dialects, at least in the pre-obstruent context (__C). Below, it will be shown that our assumption concerning the delinking of {H}

|| 44 Recall that this is one of the reasons why Brockhaus (1995) proposes that fortis obstruents in German contain {H} which is never lost, contrasting with the ‘voiced’ obstruents which lose {L} in FOD.

56 | Cracow-Poznań sandhi voicing

also in the FOD context might be correct, if not necessary for the understanding of the puzzling phenomenon of Cracow-Poznań sandhi voice assimilation, to which we turn now.

3 Cracow-Poznań sandhi voicing It was mentioned in Chapter 1 that the voicing agreement between two adjacent obstruents in Polish concerns not only the word-internal situation but also the context where two obstruents become adjacent across the word boundary. However, this simple observation is true only of the Warsaw Polish dialect. In Cracow-Poznań, the voicing agreement in the sandhi context is extended to any sequence of word-final obstruent followed by a voiced segment, including vowels and sonorant consonants. Below, we repeat the data from Chapter 1 for convenience. It should be stressed that the superscripted ‘+v’ or ‘–v’ in the contexts below are phonetic labels rather than phonological categories, and they merely serve the purpose of illustration. It will be recalled that in vowels and sonorant consonants the voicing is spontaneous in both Laryngeal Realism and Laryngeal Relativism. As for the voicing in obstruents, it is strictly dependent on the type of phonological representation. (12)

WP

CP

a. jak oni ‘how they’ wkład odrębny ‘separate contribution’

k- t-

g- __V+v d-

b. jak możesz ‘how can you’ wkład mój ‘my contribution’

k-m t-m

g-m d-m

__S+v

c. jak dobrze ‘how well’ wkład własny ‘own contribution’

g-d d-v

g-d d-v

__C+v

d. jak trudno ‘how hard’ wkład stały ‘permanent contribution’

k-t t-s

k-t t-s

__C–v

It will be recalled that in absolute final position, the obstruents are devoiced in both dialect groups. We are interested in the behaviour of the obstruent which is final in the first word with respect to the following context provided by sandhi. We are going to refer to this obstruent as the ‘target’ of sandhi assimilations, while the segment beginning the following word will be referred to as the ‘trigger’.

Cracow-Poznań sandhi voicing | 57

Beginning with the effects which are common to both dialect groups, we note that in the pre-obstruent context, the word-final target agrees in voicing with the obstruent that begins the following word (12c, d). Here, the situation looks identical to word-internal cases of the so called Regressive Voice Assimilation, which we discussed in the previous section. Thus, jak dobrze [jag db] ‘how well’ (12c), and wkład stały [fkwat staw] ‘permanent contribution’ (12d) are instances where the target obstruents do not lexically agree in voicing with the following obstruents, but they become identical with the trigger. In this sense, these forms resemble cases of dynamic assimilation, with parallels such as prośba [pr ba] ‘request’ and wsi [fi] ‘village, gen.sg.’, respectively. These facts do not require any additional machinery to be explained and theoretical models usually find them unproblematic regardless of whether they employ binary or privative melodic primes in their description. One complication that needs to be mentioned here, however, is that one should determine the role of the word boundary in such derivations. For example, in wkład stały, the surface effect of RVA may have a slightly different causality than in its wordinternal parallel wsi [fi] < /vi/ ‘village, gen.sg.’. While in the latter, the delaryngealization of the target is due to the following obstruent (vL → vo / __C), in the former, the delaryngealization of /dL/ takes place in the word-final context (__#). However, the net result is the same: the target is neutral. The rest of the derivation will depend on the nature of phonological representation, that is, on the specification or non-specification of the trigger.45 When we turn to the data in (12a, b), the difference between the two dialect groups becomes apparent. In WP, the final obstruent, just as in the absolute final position, remains voiceless in front of words beginning with vowels and sonorant consonants. To put it differently, sonorants do not voice the target obstruent in sandhi contexts. In CP, on the other hand, the word-final obstruent is voiced if the following word begins with a voiced sound. This concerns not only obstruents (12c), but also sonorants (12a, b). Additionally, this phenomenon takes place regardless of whether the word-final obstruent is lexically voiced or voiceless. If we look at the data in (12) from the point of view of voicing of the target, we observe the following pattern. In WP the target is voiced only in one context: before another voiced obstruent. Otherwise, it is regularly voiceless. On the other hand, in CP, the target is voiceless only in one context: before a voiceless obstruent. Otherwise, it is regularly voiced. To put it differently, in WP, voiced

|| 45 For the sake of argument, we assume the WP representation of the voiced fricative, that is, one with the element {L}.

58 | Cracow-Poznań sandhi voicing

obstruents are opposed to any other types of segments, while in CP, it is the voiceless obstruents that stand out against the rest. Before we look at some existing analyses of the CP sandhi voice assimilation, by which we understand the voicing of the target in front of any voiced segment, let us identify the theoretical problems concerning the CP facts. First of all, we need to understand the nature of the target itself, which is directly connected with the context in which it is found, namely, the end of the word. The data strongly suggest that the target obstruent must be deprived of any laryngeal distinctions which might be present at the lexical level. This can be concluded on the basis of the fact that both lexically voiced and voiceless obstruents are treated uniformly in the sandhi context. This uniformity takes a different form in the two dialect groups. In WP both types of obstruents end up voiceless, while in CP they are both voiced (12a, b). Thus, for example, whatever the nature of the CP sandhi voicing is, the laryngeal contrast present in the lexical or phonological representation must be first neutralized. This conclusion will be shown to be accepted in all previous analyses of the phenomenon in question.46 It will be recalled that, in the previous sections, it was proposed that CP, which is an H-system, has the same process of delaryngealization in word-final positions as WP, which is an L-system. Some support for the existence of delaryngealization in CP is provided by the parallel analysis of RVA, which was demonstrated in the previous section. However, the relevant context there was ‘pre-obstruent’ position (__C), rather than ‘word-final’ (__#). Thus, the role of the word-final context seems to be clear. This is where the delaryngealized target is produced. The remaining part of the context, namely, the following voiced sound, raises the more serious question of the nature of the trigger of the assimilation. Since, the target obstruent is voiced in equal measure before a voiced obstruent and before vowels and sonorant consonants in CP, it is tempting to assume that all these voiced segments form a natural class of assimilation triggers. This, in turn, has led some if not most analysts to the conclusion that sonorants are laryngeally specified with an active category. Only then can the phenomenon be given a phonological (computational) account.

|| 46 That the sandhi voicing in front of sonorants must involve a neutralized target was noticed quite early in the literature. For example, Baudouin de Courtenay (1894), in the discussion of the similarities between the sandhi in Polish and Sanskrit, referred to this target obstruent as ‘lenis’ (see also Vennemann 1974 for similar claims concerning Sanskrit).

Cracow-Poznań sandhi voicing | 59

It goes without saying that such laryngeal specification of sonorants contradicts the main assumptions of the privative model used here. It will be shown below that the assumption that every active-looking assimilation is due to phonological computation is in fact wrong, and the CP sandhi voicing can be given a neat explanation within the theoretical assumptions of Laryngeal Relativism. Before we provide the analysis, let us compare a few attempts at accounting for the phenomenon in question, paying attention to the target, the context and the triggers of assimilation, as well as to the differences in the rule component that have to be assumed in order to express the difference between the two dialect groups, under the assumption that representationally, we are dealing with identical systems.

3.1 Bethin (1984, 1992) The first non-linear analysis of CP sandhi voicing can be found in Bethin (1984). Bethin sets up a number of intuitively correct objectives. One of them is that the external sandhi assimilations should form an extension of the generalizations found word-internally.47 Admittedly, given the fact that CP voicing is restricted to the word-final context, it is not an easy task to account for this phenomenon as an extension of word-internal voicing assimilation rules in Polish, rather than by means of a separate rule, which would be found only in Cracow-Poznań Polish. Recall that word-internal sonorants do not affect preceding obstruents, e.g., wiosna [vjsna] ‘spring’. For Bethin, as for other analysts of Polish, the difference between /sn/ > [sn] in wiosna, and /s#n/ > [zn] in, e.g., los narodu ‘the fate of the nation’ lies in the syllabically (representationally) defined targets of assimilation. The syllable structure is crucial in defining the context in which an obstruent is a target for voicing assimilations in both dialects of Polish. In both dialects the obstruent that is the target of voicing assimilation is in the syllable appendix. Bethin makes a distinction between the syllable coda, which can only be occupied by sonorants in Polish, and syllable appendix, which is occupied by those obstruents that cannot be syllabified in the onset position. Thus, while in wiosna the word-internal obstruent–sonorant cluster can be syllabified as a complex onset, and therefore does not become the target of assimilation, in los narodu, the

|| 47 The idea that external sandhi be handled by the same computational system as wordinternal processes is also advocated in Scheer (2011: 683–695).

60 | Cracow-Poznań sandhi voicing word-final obstruent /s/ is in the appendix.48 Thus, word-internally, sonorants will never voice the preceding obstruents because these obstruents never find themselves in the appendix. In this sense, although the result is correct, we can hardly say the word-medial and the external sandhi context come from the same generalization concerning voice assimilation because the mutually exclusive contexts do not allow us to see the uniformity in action. This, however, is not an argument against Bethin’s analysis. In the analysis couched within Laryngeal Relativism, to be presented later below, the intuition that the external sandhi voice assimilation is an extension of the word-medial situation will be given a new, and more direct expression. Representationally, CP and WP are identical, and given the format of the voicing assimilation rules that Bethin proposes – for example, the use of the ‘α’ variable for voice – it may be deduced that she adheres to a binary representation. The difference between WP and CP lies in the formulation of the voicing assimilation rules. In WP, the rule is more general in the description of the target, which is simply defined as [−sonorant], and more specific in the definition of the trigger. The latter also contains [−sonorant] in order to prevent sandhi voicing before sonorants in that dialect (13a). The CP rule shows the opposite situation: the target has additional specification to do with its syllabic status, while the trigger is made very general. Any voiced segment affects the target. (13) a. WP Voicing Assimilation α voice [–sonorant] → [α voice] / ___ –sonorant

b. CP Appendix Voicing Assimilation App | C → [α voice] / ___ [α voice] [–sonorant]

|| 48 Word-medially, an obstruent ends up in the appendix only before other obstruents, not before sonorants. This is reminiscent of the difference between the contexts in which the laryngeal contrasts are sustained in Polish (/...C(S)V.../), and the ones in which the contrasts are lost (/...CC.../, /...C#/).

Cracow-Poznań sandhi voicing | 61

It is not clear why the target in (13a) is not defined in the same way as in (13b). Word-internal assimilation in WP can also take place only if the obstruent is in the appendix, which is guaranteed in pre-obstruent contexts. Clearly, the difference between WP and CP lies solely in the respective definitions of the triggers in (13). As could be predicted for an analysis that does not make a representational distinction between the two dialects, the relevant difference is expressed in the format of the phonological rule, that is, in the computational aspect. Note that, in WP, this rule accounts for both word-internal and external sandhi effects: sonorants do not voice the preceding obstruent in either context. However, given that sonorants have the feature [+voice], or receive it at some point in the derivation, the analysis offers no explanation as to why that feature does not spread to the preceding obstruent in the sandhi context, where this consonant is in the appendix, and therefore delaryngealized. To be fair, no such explanation is required in the word-medial contexts, where obstruents in front of sonorants are syllabified in onsets and retain their laryngeal specification. The explanation of the difference between obstruents and sonorants as triggers of voice assimilation is needed only for the external sandhi context. However, placing the distinction in the format of the rule appears to have a descriptive, but not an explanatory value. Bethin’s (1984) analysis has a number of advantages. For example, it avoids the use of boundary markers in phonological rules by defining the context for CP voicing representationally, that is, by means of syllabification. It is the wordfinal, not the word-internal pre-sonorant obstruents that end up in the appendix. Secondly, the intuition that internal and external assimilations should be given a uniform account, although not fully realized, certainly goes in the right direction if we want to fully understand the CP voicing. This lead will be taken up in our own analysis below. It should be added that given the precise definition of the context in which an obstruent is a target of assimilation – the appendix – Bethin’s (1984) analysis of CP sandhi voicing could also work under the privative representation of voice. However, we would have to assume that the target obstruent is not just neutral (or neutralized) with respect to its laryngeal specification, it must also belong to the appendix. The latter condition is crucial in order to distinguish between the word-internal unspecified obstruents, e.g., /so/ in wiosna [vjsna] ‘spring’, in which the fricative would be immune to pre-sonorant voicing because it is in the onset, and an identical unspecified /so/ in los narodu [lz nardu] ‘the fate of the nation’, where the fricative is in the appendix, and therefore voiced. However, the remaining crucial ingredient of such an analysis would have to be the specification of sonorants with [+voice], at least in the

62 | Cracow-Poznań sandhi voicing

Cracow-Poznań dialect. This is a move which is incompatible with the strict understanding of privativity but which we advocate in this work. Bethin (1992) attempts to extend the 1984 analysis by including privativity and modifying the role of syllabification, which is still central. The notion of syllable appendix is replaced with a positive condition concerning the relation between syllable membership and the realization as well as spreading of voice (p. 167). Thus, the target of assimilation is now defined indirectly as not being in the onset, because the laryngeal specification of obstruents is only preserved in onsets, as illustrated in (14). (14)

N’’ N X | root [–son] [+cons] | [voice]

If an obstruent is not associated to the syllable onset, its association to the laryngeal tier is lost. Consequently, only onsets show voice contrast, and trigger voice assimilation. The two contexts in which the condition in (14) is not met include, predictably, the word-final and pre-obstruent position. However, in this analysis, the privativity of the laryngeal distinction holds only at the lexical level. The generative view that phonological derivation brings underlying representations closer to the phonetic representation forces Bethin (1992) to assume that both the unspecified obstruents and sonorants receive their full specification during the course of the derivation. Bethin follows Mascaró (1987) in assuming that the value [–voice] on the unspecified obstruents is filled in at an earlier stage than the value [+voice] on sonorants, thus avoiding word-internal pre-sonorant voicing.49 The specification of the sonorants is necessary, as can be predicted, to account for the CP sandhi voicing, which for Bethin, as for most other analysts, requires a computational account. We have seen earlier, that in truly privative accounts, especially in those adhering to non-specification rather than underspecification, the presence of an assimilation in surface terms does not automatically suggest that a phonologi-

|| 49 By the time sonorants become [+voice], the potential targets of assimilation will have received their specification with [–voice], which removes them from the pool of targets.

Cracow-Poznań sandhi voicing | 63

cal assimilation is also involved. Thus, for example, the alternation koza / kózka [kza ~ kuska] ‘goat, nom.sg./ dim.’, in an L-system like WP, is a case of delaryngealization of the lexically voiced fricative (/zL/ → /zo/ > [s]), rather than a case of [–voice] spreading. Likewise, the fact that, in CP, the word-final obstruents are voiced in front of sonorants, does not have to mean that this is a case of spreading [+voice] from the sonorants, as long as a viable alternative can be given. Another problem with this improved analysis of Polish voicing is that now it is not exactly clear what defines the obstruents which are targets of assimilation. Recall that in Bethin (1984), obstruents were either [+voice] or [–voice] phonologically, as long as they remained syllabified as onsets, that is, they did not end up in the appendix. Because then the obstruents lost their laryngeal specification and were subject to assimilation via spreading. In a privative approach to voicing taken in Bethin (1992), one series of obstruents is lexically devoid of laryngeal specification, even if they are syllabified as onsets. If nothing is done about it, this would wrongly predict that such obstruents should be voiced before vowels and sonorant consonants inside words. Thus, to avoid the unwelcome result of spreading [+voice] onto such obstruents inside words, e.g., in wiosna, they must be filled with [–voice] before the sonorants receive their [+voice]. This order of events is compatible with the stages of laryngeal specification proposed in Mascaró (1987). However, in order to get the right results in Cracow-Poznań sandhi assimilations, the order in which the final obstruents and initial sonorants receive their laryngeal specification must be reversed. Otherwise, such assimilation would be impossible. In models resorting to rule ordering, this problem is solvable. It is enough to say that there are two bouts of voice filling. One of them concerns onsets, and allows the unspecified obstruents to receive [–voice] before sonorants get [+voice]. The other bout would concern word-final obstruents, which are not in the onset. This default filling must follow the assignment of [+voice] on sonorants. The problem is that this way we are getting further away from the intuitively correct observation that external sandhi voice assimilations should constitute an extension of the wordinternal situation, not to mention the complexity of the analysis involving rule ordering. To account for WP, in which no pre-sonorant sandhi voicing occurs, Bethin allows only the CP sonorants to spread [voice]. The spreading rule in (15), which is proposed in Bethin (1992: 184), is meant to be general and encompass both word-internal and the external sandhi assimilations in CP.

64 | Cracow-Poznań sandhi voicing

(15)

Cracow-Poznań Voice Spread X1 | root [+cons][–son]

X2 | root | [voice]

Bethin suggests that ‘the only modification necessary is to note that in the CP dialect voicing may spread to obstruents from following sonorants at the phrase level’ (Bethin 1992: 184). Unfortunately, although descriptively correct, the notion ‘at the phrase level’ is not expressed anywhere in the above rule, leading to the theoretical problems with ordering of the voice specification rules inside words as suggested in our discussion above (14). Thus it seems that while adopting the privative representation of voice is a good move – even though this type of privativity is unrestricted in the sense that all unspecified values must at some stage be filled in – the abandonment of the syllable appendix as determining the assimilation target renders this analysis less appealing. However, a less general formulation of the CP Voice Spread rule, that is, one in which syllabic affiliation of the target obstruents would be made more specific, would only improve things slightly. It would completely divorce the external sandhi assimilations from the word-internal ones, still leaving the problem of targets unresolved. If there is such a process as [voice] spreading from sonorants to neutral obstruents, it should be able to affect both the lexically unspecified series, and those which have lost the lexical specification, with no regard for the actual syllabic configuration. Below, we look at two analyses couched in a binary model.

3.2 Gussmann (1992) Gussmann (1992) uses a binary feature system, in which the voiced obstruents carry [+voice], and the voiceless obstruents have [–voice]. Sonorants are assumed to be underspecified, but, as in Bethin (1992), they become [+voice] in the derivation. Exactly when this filling takes place and by what mechanism – presumably some default – is not mentioned. In this analysis, voicing must be licensed both in sonorants and in obstruents, and the licensing in obstruents takes place only if they are syllabified as onsets, while the sonorants license voice in onset or coda position. The sonorants may lose the laryngeal specifica-

Cracow-Poznań sandhi voicing | 65

tion if they are not syllabified. This concerns such forms as mędrka [mntrka] ‘smart aleck, gen.sg.’, and wiatr [vjatr] ‘wind’. The account of FOD takes two steps. Firstly, the word-final obstruent, which is not syllabified in the onset, loses its voicing category. Such obstruents are then filled with [−voice] by default, unless [+voice] is spread from the following context. On the other hand, word-internal Voice Assimilation as in liczba [ljidba] < /lit ba/ ‘number’, takes the form of a rule of Obstruent-to-Obstruent spreading (Gussmann 1992: 44). We show it below with some modification. (16)

Obstruent-to-Obstruent spreading α voice

[+obstruent]

[+obstruent]

For the word-external Voice Assimilation in Warsaw Polish, Gussmann proposes a more general rule with respect to the target, in that this type of assimilation is due to the rule of Obstruent-to-Consonant spreading, where ‘consonant’ stands for both obstruents and sonorants. This extension is needed to account for the fact that, for example, in wiatr zachodni [vjadr zaxdi] ‘west wind’, the voice spreading affects both the obstruent and the sonorant, which is otherwise voiceless in wiatr [vjatr] ‘wind’. It is interesting to note that, unlike for Bethin (1984, 1992), the fact that the rule of word-external Voice Assimilation is an extension (generalization) of the one operating word-internally is taken by Gussmann as an argument for treating the two types of assimilation as different. This is clearly reflected in his formalization of the CP sandhi voicing rule, in which the target remains general. It is simply a ‘consonant’. Additionally, the trigger is now extended to all types of voiced segments, that is, obstruents, sonorants and vowels. (17)

CP sandhi voicing +voice

[+consonant]

X

(X = any segment)

This extension of the types of triggers of CP sandhi voicing is similar to what we observed in Bethin (1984, 1992). It will be seen presently that it is also no different from the proposal in Rubach (1996). In our view, however, an analysis which

66 | Cracow-Poznań sandhi voicing

uses such extension in the rules of assimilation in order to account for the CP voicing in fact admits defeat. While descriptively correct, such analyses have no explanatory value. It is still unexplained why in this particular dialect sonorants should voice obstruents, and why only across word boundaries.

3.3 Rubach (1996) The analysis of Polish voicing presented in Rubach (1996) appears to go against the main theoretical trends of the time. However, it is the most comprehensive and workable account to date. The binary view of features, and especially rule ordering, which lies at the heart of Rubach’s analysis, is indeed contested in the literature. The early and mid nineties abound in arguments in favour of privativity, especially with respect to laryngeal phonology (e.g., Avery 1996; Bethin 1992; Brockhaus 1995; Harris 1994; Iverson and Salmons 1995; Lombardi 1991, 1995b), while rule ordering has by and large been given up in mainstream phonological theory. The author goes as far as to admit himself that his two main assumptions, namely, that [voice] is binary rather than privative, and that voice assimilations are not syllable-based, are highly controversial (p. 76). In fact, his decision not to treat Polish voicing phenomena as stemming directly from syllabification is very much in keeping with the linguistic trends. The syllable was denied existence in a number of approaches of the early 1990s (e.g., Kaye, Lowenstamm and Vergnaud 1990; Dziubalska-Kołaczyk 1995). Like Bethin and Gussmann, Rubach assumes that the voicing specification of sonorants is related to prosodification, and that at some point in the derivation, sonorants become [+voice] as per Sonorant Default. Another similarity to the previous analyses concerns the analysis of FOD. As in Gussmann (1992), it is a two-step phenomenon. The first one is delaryngealization at the right edge of the prosodic word, and the second step consists in the application of Voice Default, a fill-in rule which supplies [−voice] to segments which are unspecified for [voice] (p. 77). Since, lexically, both series of obstruents carry a laryngeal feature, both types are delaryngealized in word-final position. In other words, in that position they can have neither [+voice] nor [–voice]. In this sense, the delaryngealization produces ‘a third object’, which has no laryngeal node – something similar to what we symbolize as /Co/ in our discussion. This neutralized obstruent must be subject to further derivation. It receives either [+voice] or [– voice] by spreading from other local objects that possess these properties, or it is subject to default filling. In other words, the delaryngealized object never surfaces as such.

Cracow-Poznań sandhi voicing | 67

Word-internal voicing assimilation, in Rubach’s binary representation of voice, always involves familiar delaryngealization of an obstruent in front of another obstruent, followed by leftward spreading of the laryngeal node from the trigger (p. 78). Since the following obstruent is specified with either [+voice] or [–voice], it appears that such a binary system is suitable for modelling the symmetrical assimilations of voice which we observe in Polish. As for sandhi assimilations before words beginning with obstruents, they work in a similar way. What is different is the cause of delaryngealization of the target consonant (word-final obstruent). It is deprived of the laryngeal node due to its position at the end of the prosodic word. Thus, for example, in Warsaw Polish, where only voiced obstruents may assimilate the preceding word-final obstruents, e.g., sad wiśniowy [sad vjiv] ‘cherry orchard’, the presence of the assimilation is attributed to the regular Obstruent-to-Obstruent Spread rule, which is identical to that operating word-internally.50 Cracow sandhi voicing, on the other hand, is given a separate rule called Cracow Spread (p. 82), which can be described as spreading of the laryngeal node from any root node (of an obstruent, sonorant or vowel) to an adjacent root node lacking a laryngeal specification. (18)

Cracow Spread R

R

Lar In principle, then, this analysis of CP voicing is very much the same as those proposed by Bethin (1984, 1992) and Gussmann (1992). One of the conditions on the rule of Cracow Spread is that the triggers are prosodified and hence already filled with [+voice] – by Sonorant Default in the case of vowels and sonorants. The filling must crucially occur before the operation of this rule. On the other hand, the target must not be filled with [–voice] prior to the application of this rule. It is interesting to note the generality of the formulation of this rule. No word boundary is mentioned, and the rule can easily be extended to operate also word-medially in that particular dialect, which responds positively to the postulate formulated already in Bethin (1984) that internal and external assimilations be treated uniformly. Indeed, Rubach can || 50 In front of voiceless obstruents, we have two options. Either there is [–voice] spreading, or the word-final obstruent receives this feature by default.

68 | Cracow-Poznań sandhi voicing

claim that this is the case, except that, as in Bethin’s proposal, Cracow Spread never gets the chance to operate word-internally because the configuration in (18) never occurs in that context. This is due to two assumptions. Firstly, [voice] is binary in Polish, and secondly, there is no delaryngealization of obstruents in front of sonorants. Thus, the rule in (18) is in effect limited to sandhi – CP sandhi, to be precise. The rule in (18) appears to work best in a binary system, and it would be problematic if combined with privativity, as understood and applied in, for example, Bethin (1992). If sonorants in Cracow-Poznań Polish spread [+voice] onto preceding obstruents with no laryngeal specification, they should do so also word-medially. This flaw of the privative analyses, which was mentioned above as well, was scrupulously noted in Rubach (1996). It would seem that to salvage privativity, one should return to the presence of the word boundary in the phonological rule of CP voicing, something that all the above mentioned authors wanted to eliminate, or mark the targets of CP voicing more precisely, for example, the way Bethin (1984) did, that is, by reference to such representational devices as the appendix. However, there is a much more fundamental problem with the rule in (18), as well as with the privativity assumptions used in Bethin (1992). Namely, strictly privative approaches like Laryngeal Realism or Laryngeal Relativism cannot allow sonorants to possess laryngeal features. Hence, there is nothing to spread, and rules of the type given in (13b), (15), (17), and (18) above, are simply impossible. The question is if an explanatory analysis of CP sandhi voicing is possible without referring to computation, that is, to spreading of a laryngeal category from a sonorant onto the target. Below, we demonstrate that such an analysis is possible if Laryngeal Realism is replaced with Laryngeal Relativism.

3.4 CP sandhi voicing and Laryngeal Relativism It was demonstrated in Chapter 1 that CP sandhi voicing is inexplicable within Laryngeal Realism (Harris 1994; Honeybone 2002). A fairly recent and comprehensive analysis of the Polish voicing complex within that model, which can be found in Gussmann (2007), ground to a halt with respect to two phenomena in modern Polish: CP sandhi voice assimilation, and Progressive Voice Assimilation. The latter phenomenon is discussed at length in the following chapter. Gussmann (2007) assumed that Polish is an L-system regardless of the dialectal division. In this type of system, an obstruent may only be voiced if it has or receives the element {L}. The non-specified obstruents (/Co/) must be phonetically realized as voiceless unaspirated. At the same time, sonorants are not

Cracow-Poznań sandhi voicing | 69

allowed to carry a laryngeal element. It follows that in such a representationally restrictive model, the CP sandhi facts simply cannot be modeled. This is because obstruent voicing in such a system must be truly phonological. Thus, the combination of the non-specification of sonorants and an L-system, in which obstruent voicing must be due to representation or computation involving {L}, predicts the absence of pre-sonorant voicing in WP. For this reason, we assume that an L-system has its place in the Polish voicing complex, albeit only in Warsaw Polish. We proposed that Cracow-Poznań Polish is a different system, which uses the {H} element in its voiceless unaspirated obstruents, while the voiced obstruents are phonologically non-specified. Their voicing is interpretational. We refer to it as enhanced passive voicing, which requires phonetically voiced environment, like any passive voicing in obstruents. Additionally, in the above sections, we have demonstrated that this ‘reversed’ system is equally capable of accounting for the same major phenomena to do with voicing in Polish, that is, FOD and RVA. Before we look at a comparison between WP and CP in sandhi contexts, let us remind ourselves of some points which we have already established with respect to the targets and triggers of voicing in both dialects. Earlier in this chapter, it was demonstrated that the word-final obstruents in WP are delaryngealized in order to account for FOD. Although the so called final devoicing in CP takes a different form – it is interpretational rather than computational – we assumed that the element {H} is also delinked in this dialect, not only in the word-final context, but also word-medially in front of an obstruent. The latter context yields RVA facts. The assumption that {H} is also lost word-finally, finds strong support in the fact that the targets of external sandhi assimilation do not exhibit a difference based on their lexical representation of voice. Namely, both marked and unmarked obstruents behave uniformly in external sandhi in both dialects. Thus the phonological definition of the target in both dialects is that it is laryngeally non-specified. (19)

Word-final target WP/CP = /Co/#

While representationally the target is identical in both dialects, it belongs to two disparate systems in which the conditions on voicing are different. In WP, the obstruent must retain its lexically present {L} or receive this element via spreading in order to be pronounced as voiced. In CP, on the other hand, the voicing

70 | Cracow-Poznań sandhi voicing

requires a phonetically adjacent voicing of another segment. It does not require a phonological category. (20)

Triggers of voicing in obstruents WP= {L} C

L

CP = voiced context Co

This is what we demonstrated above in the sections devoted to FOD and RVA. It follows that a neutral obstruent in the word-final context will behave in exactly the same manner as it does word-medially. Note that this means that the postulate that word-internal and word-external assimilations should be treated in the same way (Bethin 1984), finds a full expression in this model. Word-internally, the non-specified obstruents in CP are voiced in front of vowels, sonorant consonants and other non-specified obstruents in voiced environment. This is exactly what we find in the so called CP sandhi voicing. (21)

Voiced obstruents in CP word-internal51

sandhi context52

CoV [dm] CoS [brat] CoCo [gd]

Co#V [brad-] Co#S [kub-rb] Co#Co [jag-db]

= = =

Note that in CP, vowels, sonorant consonants and neutral obstruents indeed form a kind of natural class, but not in the usual sense of the word. What makes them similar is that their voicing is not due to an active laryngeal category: they are all non-specified for voice. We see that there is, in principle, nothing wrong with pre-sonorant voicing, but it must occur in a particular system in which the phonetic interpretation of non-specified obstruents takes the form of enhanced passive voicing and is contextually conditioned. The conditioning must be the same for word-medial and word-external voicing. The conditions that must be fulfilled for such effects to occur are the following. Firstly, we must be dealing with an H-system, in which for some reason the phonetic interpretation of the marked series does not show aspiration. Only then do we get enhanced passive voicing of the unmarked series. Secondly, the sandhi pre-sonorant voicing is

|| 51 dom ‘house’, brać ‘to take’, gdy ‘when’. 52 brat ojca ‘father’s brother’, kup rybę ‘buy fish’, jak dobrze ‘how well’.

Cracow-Poznań sandhi voicing | 71

possible only with respect to the unmarked target. And finally, there must be phonetic adjacency between the target and the so called trigger. Below, we compare the two dialects in sandhi contexts in more detail, looking only at the relevant parts of representation in a prevocalic context. Note that word-finally both WP and CP allow for non-specified obstruents only. Either it is a lexically non-specified object (/Co/), or it is a neutralized one. The neutralization or delaryngealization takes the form of L-delinking in WP (/CL#/ → /Co#/), and H-delinking in CP (/CH#/ → /Co#/). If the word is pronounced in isolation the resulting non-specified object is realized as voiceless in both dialects. On the other hand, when the following word begins with a vowel, which provides phonetic, not phonological adjacency, the neutral obstruent in WP must still be realized as voiceless (22a, b) because the delinking of {L} cannot be undone, while in CP, it must be realized as voiced (22c, d). (22)

Warsaw Polish a. jak oni ‘how they’ Co # V | | j k  i

b. wkład odrębny ‘separate contribution’ L-delinking → Co # V CL # | | | fkwad fkwad d r m b n 

phonetic adjacency and interpretation

ko> [k]

do > [t]

Cracow-Poznań Polish c. jak oni ‘how they’ H-delinking → Co # V CH # | | | ja k jak  i

ko> [g]

d. wkład odrębny ‘separate contribution’ Co # V | | fkwad d r m b n 

do > [d]

72 | Cracow-Poznań sandhi voicing

Under this analysis, Cracow sandhi voicing is not a result of an arbitrary rule, as previous accounts of the phenomenon would have it. It is not even a case of phonological computation. It is simply an obligatory interpretation, given the laryngeal system of that dialect. The Laryngeal Relativism view allows us to understand Cracow sandhi voicing in the following way. Phonology is responsible only for the reversed representation of the voicing contrast and the word-final delaryngealization producing /Co/. The sandhi voicing is not due to phonological spreading of a laryngeal feature, which would force us to assume that sonorants carry laryngeal categories, but due to interpretation of a neutral object which is phonetically adjacent to a phonetically voiced segment, something that is also regular word-internally in that dialect. The main features of this analysis that we find advantageous are the following. A strict privative representation of the laryngeal contrasts is used, which is similar to Laryngeal Realism in a number of aspects. Firstly, the non-specified objects are not further derived by phonological computation in order to be pronounceable. They are simply interpreted according to the systemic phonetic interpretation conventions, which, for example, allow for a fully voiced interpretation of the unmarked series in an H-system, if the marked one is not aspirated. This enhanced passive voicing requires a phonetically voiced context. Secondly, sonorants remain non-specified with respect to voicing, and their participation in voice assimilations is restricted to a system and contexts in which this phenomenon may take a purely phonetic and interpretational form. This is the case in Cracow-Poznań Polish. Thirdly, the analysis of CP sandhi voicing does not rely on any rule ordering. If there is any ordering of events we must stipulate that phonological computation precedes phonetic interpretation. This is hardly a questionable assumption. Finally, no ad hoc rule of voice assimilation is required which designates sonorants and vowels in one dialect to spread voicing. Given the two types of laryngeal systems in WP and CP, the facts in both dialects come out directly from the design of the systems in question. It will be recalled that one of the objections to privative laryngeal models of representation was that sonorants which voice word-final non-specified (delaryngealized) obstruents should do the same word-medially, ostensibly leading to wrong forms like *[zzna] instead of [ssna] for sosna ‘pine tree’ (e.g., Rubach 1996: 108). We are now in a position to say that this argument was correct, but it could not be used as an objection for dialects like CP, given that there such ‘voicing’ is in fact expected in word-medial contexts, but not in forms like /sHsHna/ > [ssna], in which the element {H} finds itself in ‘friendly’ contexts (/__V/, /__(S)V/), but in forms like pozna [pzna] < /pHzona/ ‘will meet’, in

Cracow-Poznań sandhi voicing | 73

which the non-specified fricative is indeed ‘voiced’ due to the phonetic context in which it occurs. Now that we have distinguished the two types of laryngeal systems in Polish – that is, the H-system of CP, and the L-system of WP – the objection cannot hold also for the Warsaw dialect any more. Since in sandhi, sonorants do not voice the preceding obstruents in that dialect, why should we expect them to do so word-medially? Thus the reversed WP representations /pozLna/ and /sosona/ are still expected to yield the same results, that is, [pzna] and [ssna], respectively.

3.5 Laryngeal Relativism and Licensing-by-Cue Finally, a comment is necessary with respect to another alternative proposal to deal with laryngeal phonology in general and pre-sonorant voicing of the Cracow-Poznań type in particular. So far, in our discussion, we have left out one important aspect concerning the distribution of laryngeal contrasts and the contexts for neutralization. To be more precise, we have identified the contexts for delaryngealization in Polish but without giving a particular theoretical explanation as to why word-finally and in pre-obstruent position the contrast – in our case, element {H} or {L}, depending on the dialect – is lost. However, the choice of a phonological model in our discussion, that is, a version of Government Phonology, makes it clear that the distribution of the laryngeal elements {L} and {H} in Polish will be controlled by a form of licensing. Thus, our approach belongs to the prosodic licensing tradition (e.g., Goldsmith 1990; Itô 1986; Lombardi 1991, 1995a), but, it avoids, hopefully, all the pitfalls of syllablebased analyses that were identified, for example, by Rubach (1996), Steriade (1999), and Strycharczuk (2012b).53 The prosodic licensing approach has a strong competitor in the form of the Licensing-by-Cue hypothesis put forward in Steriade (1999). The essence of licensing by cue is that laryngeal neutralization tends to occur in phonetic contexts in which some, or most cues to do with the laryngeal opposition is missing. For example, while intervocalically (V1CV2) there are seven possible cues to voicing (closure voicing, closure duration, V1 duration, F1 values in V1, burst duration and amplitude, VOT value, F0 and F1 values at the onset of voicing in V2), word-finally and in pre-obstruent position the number is reduced to five or

|| 53 The discussion of the precise conditioning of the distribution of elements {L} and {H} in Polish is offered in Chapter 5.

74 | Cracow-Poznań sandhi voicing

four respectively, with the most robust cues being lost (Steriade 1999: 31). Thus, Steriade’s model is potentially very precise in providing the contexts in which neutralization is expected more than in others. The question is how this phonetic information is reflected in the grammar, or in fact in phonology. Steriade does not make it clear how licensing itself could be formalized, but clearly it is not part of phonological representation. It is not our intention to compare our approach to Steriade’s with respect to the distribution of laryngeal features. This will be done to some extent in Chapter 5, where we will attempt to show that, contrary to Steriade’s claims, prosodic licensing can provide a uniform description of neutralization contexts. They just cannot be syllable-based. What we would like to concentrate on now is Steriade’s account of pre-sonorant voicing in external sandhi contexts, or its absence. This comparison can be made successfully without a precise reference to delaryngealization because this aspect of the analysis of CP sandhi voicing in the two approaches is identical: we both assume delaryngealization as one of the conditions. To be more precise, Steriade does not discuss CP sandhi directly. However, she provides a hypothetical solution that can be offered for similar if not identical phenomena in Sanskrit (Steriade 1999: 82). But first let us mention how she would handle data of the Warsaw Polish type, in which no pre-sonorant sandhi takes place. There is also no direct reference to Warsaw Polish in this respect, but Steriade discusses a very similar system in Lithuanian, in which voiced stops are fully voiced and voiceless stops are voiceless unaspirated. In Lithuanian voicing is lost word-finally and in pre-obstruent context. Thus, the word for ‘much’, which lexically ends in a voiced obstruent, that is /daug/, is pronounced [dauk], with FOD. If the following word begins with a vowel, the neutralization of the final obstruent is still maintained, and it is pronounced as voiceless, e.g., dau[k a]kmens ‘many stones’. To account for this cyclic effect despite the presence of the phonetic context providing potential cues to voicing,54 Steriade has to appeal to the notions of morpheme invariance or paradigm uniformity. Thus she proposes that in such cases the voiceless realization of the final obstruent is due to highly ranked ‘constraints promoting morpheme invariance or minimization of allomorphy. More specifically, the suggestion is that the voice-neutralized word edge takes on invariably the form that would be phonetically natural in the citation form’ (p. 80). She also adds that ‘this type of paradigmatic extension takes place only between word-forms.’ This analysis, based on paradigm uniformity, must be viewed as extrinsic because the ranking

|| 54 It seems that cyclic effects do not sit well with phonetically based approaches.

Cracow-Poznań sandhi voicing | 75

of these constraints must be different to account for systems like CracowPoznań Polish. The use of other than phonological information in the analysis is also a drawback. Note also, that the paradigm uniformity constraints in question must be limited to the right edge of the word. Thus, compared to our analysis, in which the absence of pre-sonorant voicing in WP falls out automatically from the phonological representation of the voicing contrast and delaryngealization wordfinally, Steriade’s account of the WP type system is unduly complex. We do not need to refer to paradigm uniformity, because the delaryngealized obstruents in WP must be interpreted as voiceless. The voiced phonetic context cannot reverse the neutralization for two reasons. Firstly, it is the phonology that decides on neutralization. And secondly, voiced obstruents in WP must receive a phonological category {L} that sonorant consonants and vowels lack. Let us now turn to CP sandhi voicing. Referring to the Sanskrit cases which are parallel to CP, Steriade provides a solution based on a distinction between auditory and articulatory features. She claims, that in systems like Lithuanian (and Warsaw Polish) the paradigm uniformity constraints require auditory identity, while in languages like Sanskrit (and Cracow-Poznań Polish) the required identity is of an articulatory nature. Thus, while in WP the word-final obstruents must be voiceless, the corresponding CP ones must be lenis. It must be admitted that it is rather difficult to argue against Steriade’s proposal from within the model that she adopted. This model allows for a mixture of phonological, phonetic and extraneous paradigmatic constraints in one computational system. However, it makes an observation with which we must intuitively agree. In our analysis, the neutral obstruents in CP are lenis, because that type of articulation allows for passive voicing. For these reasons, we would like to suggest that our model and the Licensing-by-Cue hypothesis are not entirely incompatible. The structure of the laryngeal system that we proposed in the previous chapter and applied to CP sandhi voicing here, does not exclude a very complex and multi-faceted conditioning of the phonetic interpretation side. Interpretationally, speakers will do whatever is necessary in order to produce the right phonetic targets. Given that Steriade herself stresses the implementational character of the conditioning for cue distribution, we see ground for a fruitful integration. One condition that we can envisage here is that the role of phonological representation be clarified in implementational models. We saw in this chapter that phonetic interpretation in a given system is strictly dependent on the type of phonological representation, including such aspects as privativity and non-specification of sonorants. Steriade (1999) is not clear on either of these points.

76 | Cracow-Poznań sandhi voicing

4 Conclusions This chapter argued for a treatment of the voicing facts in WP and CP as stemming from two different, in fact opposite, laryngeal systems. WP is an L-system, in which the marked series of obstruents contains {L}, while the non-specified series must be phonetically realized as voiceless unaspirated in all positions. Voicing of obstruents in this system is possible only if the element {L} is present or spread from a local source. CP is an H-system, in which the marked series of obstruents contains {H}. However, the phonetic interpretation of this representation does not involve aspiration, as Laryngeal Realism would expect. Rather, these objects are pronounced as voiceless unaspirated. Due to systemic interpretation conventions, the non-specified series of obstruents are fully voiced. This voicing is called enhanced passive. It was shown that the bi-systemic analysis is able to handle the basic facts concerning Polish voicing, such as Final Obstruent Devoicing and Regressive Voice Assimilation. More importantly, if Cracow-Poznań Polish is an H-system, then the CP sandhi voice assimilations emerge directly from the systemic interpretation conventions which also hold word-internally. The consequences of this proposal are far-reaching. If the analyses are correct, then a number of classic criteria used for determining the categorical status of voicing will have to be abandoned, or at least treated with more caution henceforth. For example, long VOT lead in fully voiced plosives, coupled with the presence of voice assimilation to a voiced obstruent have often been used as diagnostic criteria for claiming that we are dealing with a ‘voicing’ language, in which the voiced series contain a laryngeal feature [voice] or element {L}, depending on a model. If CP is indeed an H-system, as we have argued in this chapter, then it fulfils the criteria to be a ‘voicing’ language, but it is not one. One may agree that it is a ‘voicing’ language in phonetic terms, but it is certainly not a ‘voicing’ language in phonological and indeed systemic terms, because phonologically, it is an ‘aspiration’ language. Another typical effect of ‘voicing’ languages, namely, FOD understood as delaryngealization, also needs to be reconsidered. A number of analyses in the existing literature have also questioned this understanding (e.g., Brockhaus 1995; Harris 2009). Final devoicing may take the form of an absence of passive voicing in a context which inhibits spontaneous vibration of the vocal cords. It is expected in systems with passive voicing, that is in ‘aspiration’ languages which are H-systems, as well as in ‘voicing’ languages which are H-systems. Thus, in a sense, all phonetic facts are systemic, but some of them have a phonological basis, and others an interpretative one. To see this clearly, one

Conclusions | 77

should have a clear idea of what a sound system is, and one should be able to distinguish between the two bases of sound patterns. In the following chapter, we take up the problem of the so called Progressive Voice Assimilation in Polish (PVA). The analysis will not so much depend on the distinction between L-systems and H-systems, as on the very nature of the relation between phonology and phonetics which follows from the proposed structure of sound systems in Chapter 2. Among other things, it will be shown that the model makes clear predictions and claims concerning phonological change.

Chapter 4 Progressive voice assimilation in Polish 1 Introduction The phenomenon referred to as Progressive Voice Assimilation (PVA) concerns two fricatives in modern Polish, which historically came from /w/ and the palatalized /r/, respectively. Synchronically speaking, they fall into the voiced [v], [] and voiceless [f], [ ] types, depending on the context, as illustrated in the following scheme. (1)

a.

b.

Original

w

rj

Later

Modern Polish

Contextual distribution i.

ii.

iii.

iv.

v

D_

_D

#_

V_V

f

T_

_T

_#



D_

_D

#_



T_

_T

_#

v

V_V

ř55

D = voiced consonant, T = voiceless consonant, V = vowel, # = word boundary This diachronic development, which is still reflected in the spelling, e.g., dwa [dva] ‘two’, grzebać [gbat] ‘bury’ as well as twój [tfuj] ‘your’, krzak [k ak] ‘bush’, is one of the reasons why progressive assimilation was noted quite early in the modern Polish linguistic tradition.56 However, the typical phonological arguments put forward in favour of treating these patterns as following from a synchronic process are based on distributional restrictions and alternations involving [v~f] and [r~/ ] pairs. || 55 This sound, which is found, e.g., in modern Czech, is described as a diphthongal realization [rz] before turning to [] (Klemensiewicz, Lehr-Spławiński and Urbańczyk 1955: 148). 56 Intervocalically, for example, both ‘w’ and ‘rz’ are always voiced [v] and [ ], respectively.

80 | Progressive voice assimilation in Polish

The above arguments are further strengthened by the fact that in a few dialectal areas the Tv sequences are pronounced with a voiced fricative, e.g., twój [tvuj] ‘your’ and kwiat [kvjat] ‘flower’.57 These forms are particularly instructive because, contrary to the overwhelming pattern of voice agreement in obstruent clusters, we are dealing here with an absence of progressive, or regressive voice assimilation, and consequently, with a violation of the general phonotactic regularity in Polish. To begin with the labio-dental fricative, it is assumed that it was a semivowel /w/ in Common Slavic (Stieber 1973: 14). Thus, historically speaking, we are dealing first with strengthening, or obstruentization of [w] to [v], allowing for further modifications based on the voicing properties (e.g., Cyran and Nilsson 1998). In modern Polish, the voiced [v] is found in typical non-devoicing contexts for obstruents, e.g., #_V, V_V, etc.), while the voiceless [f] is found in typical devoicing contexts, e.g., _#, _T, and after voiceless consonants, i.e., T_.58 The relevant data relating to the contexts i–iv in (1a) are given below. They show both static distributional restrictions (2a) and alternations (2b). (2)

a. Dv dwa [dva] ‘two’ gwiazda [gvjazda] ‘star’ gwar [gvar] ‘hubbub’ zwać [zvat] ‘call’ żwawy [vav] ‘brisk’ b. Tf

Tf twarz [tfa ] ‘face’ kwiat [kfjat] ‘flower’ kwas [kfas] ‘acid’ swatka [sfatka] ‘matchmaker’ szwaczka [ fat ka] ‘tailor’

f#

VvV

marchwi [marxfji]

marchew [marxf]

marchewek ‘carrot, gen./nom./dim.gen.pl.’ [marxvk]

Tf

f#

VvV

cerkwi [tsrkfji]

cerkiew [tsrcf]

cerkiewek ‘Orthodox church, gen./nom. [tsrcvk] /dim.gen.pl.’

listwa [ljistfa]

listw [ljistf]

listewek [ljistvk]

‘board, nom./gen.pl./dim.gen.pl.’

|| 57 The voiced fricative [v] in Tv contexts is retained in Kashubia, Great Poland and parts of Silesia (Dejna 1973: 100). 58 Since it acquired a phonemic status, [f] is also found in non-devoicing contexts in modern Polish, e.g., fala ‘wave’, trafić ‘hit the target’.

Introduction | 81

Returning now to the post-alveolar fricatives [] and [ ] which come from the historical palatalized [rj], the data are to a great extent parallel to those observed with Tv/Dv. Two differences, however, need to be mentioned here. Firstly, unlike the labio-dental fricative, there are no dialectal reflexes of a voiced postalveolar fricative after a voiceless consonant *[t, p, k], though some dialects are reported to have retained the earlier [tř, kř].59 Secondly, the evidence for progressive voice assimilation in T in the form of alternations is of a debatable and rather indirect status, as they involve [r~] or [r~ ] exchanges, rather than direct [~ ] alternations. To be more precise, the relation between [] and [ ] can be seen only through the mediation of [r] in historically non-palatalized forms, for example, góra [gura] ~ górze [gu] ‘mountain, nom./dat.’, or lekarz [lka ] ~ lekarski [lkarsci] ‘doctor / medical’. Gussmann (2007) deems such alternations morphophonological rather than phonological in nature. The data in (3a) show static distribution of voiced and voiceless post-alveolar fricatives, while (3b-f) illustrate the alternations with [r]. (3) a. D grzywa [gva] ‘mane’ drze [d] ‘he/she tears’ grze [g] ‘game, dat.’ b. VrV para [para] ‘couple’ wiara [vjara] ‘belief’ stary [star] ‘old’ c. T wytrzeć [vt t] ‘wipe’ wyprzeć [vp t] ‘force out’ d. D wydrzeć [vdt] ‘tear out’ wywrzeć [vvt] ‘exert’

T krzywa [k va] ‘uneven, fem.’ trze [t ] ‘he/she rubs’ krze [k ] ‘ice floe, dat.’ VV parze [pa] ‘couple, dat.’ wierze [vj] ‘belief, dat.’ starzy [sta] ‘old, pl.’ VrV wycierać [vtrat] ‘wipe, imperf.’ wypierać [vpjrat] ‘force out, imperf.’ VrV wydzierać [vd rat] ‘tear out, imperf.’ wywierać [vvjrat] ‘exert, imperf.’

|| 59 This concerns, for example, the area of Silesia which borders with Moravia (Klemensiewicz, Lehr-Spławiński, and Urbańczyk 1955: 149; Stieber 1946a: 77). Such forms are on the wane.

82 | Progressive voice assimilation in Polish e. T kutrze [kut ] ‘cutter, loc.’ futrze [fut ] ‘fur coat, dat.’

Tr kutrem [kutrm] ‘cutter, instr.’ futro [futr] ‘fur coat’

f. DDr dobrzy [db] ‘good, pl.’ wiadrze [vjad] ‘bucket, dat.’

dobry [dbr] ‘good, sg.’ wiadro [vjadr] ‘bucket, nom.’

For the data involving [r~/ ] to be fully parallel to those involving the labiodental fricative, one would expect to find alternations involving only the voicing aspect of the post-alveolar fricative [~ ]. In other words, we would need to have data parallel to marchwi / marchewek [marxfji ~ marxvk] ‘carrot, gen./ dim.gen.pl.’. Such data are missing in Polish, as illustrated below. (4)

T

VV

kutrze [kut ] ‘cutter, loc.’ futrze [fut ] ‘fur coat, dat.’

*kuterzem [kutm] *futerzak [futak]

In this sense, the data concerning the two types of fricatives are not fully comparable. A possible explanation of this state of affairs will be offered in the section devoted to our analysis of the Progressive Voice Assimilation. Finally, a comment is in order concerning the presence of [tv, kv] in modern Polish. It strongly suggests that what sounds phonetically as a labio-dental fricative [v] may fail to cause or undergo voice assimilation. In other words, it behaves like other sonorants, such as, [l, r, n, m, j, w]. The modern sequences [tv, kv] with a labio-dental voiced fricative also suggest that, even if historically correct, there is no need to evoke the bilabial articulation of the fricative in order to explain its behaviour in Tv.60 Such attempts may be understood as stemming from an assumption that a fricative articulation automatically means that we are dealing with an obstruent. Interestingly, while friction is both a phonetic and a phonological property of consonants, obstruency is not such a clear category for either domain of investigation. At what point does a sonorant become an obstruent phonetically or indeed phonologically? These issues will be returned to in the following sections.

|| 60 Klemensiewicz, Lehr-Spławiński, and Urbańczyk (1955: 136) assume that it is the bilabial articulation of [v] that contributed to the retention of a voiced fricative in Tv contexts in some modern Polish dialects.

Introduction | 83

To conclude the discussion above, the obvious problem concerning the behaviour of the two fricatives in Polish lies in the fact that, contrary to the regular resolution of sequences of obstruents with disparate voicing by means of Regressive Voice Assimilation (RVA), the data presented above seem to point to the opposite direction of influence. The evidence that we are dealing with Progressive Voice Assimilation (PVA), or in fact, only progressive devoicing, in, e.g., trze [t ] ‘he/she rubs’ and listwa [ljistfa] ‘board, nom.sg.’ is rather complex and involves a mixture of historical indication – the fact that historically the relevant fricatives were sonorants [w] and [r] – phonological and morphophonological facts based on voice alternations [v~f] and the alternations [r~/ ], respectively.

1.1 Benni’s generalization Benni (1909) is a short paper in which the author makes a number of interesting observations and generalizations about voice assimilation in Polish within the limit of the phonological word. The discussion does not include the external sandhi and concentrates only on the assimilations between obstruents. Benni’s aim is to find a principle or generalization which would show that progressive and regressive voice assimilations are in fact one and the same thing. The first interesting observation that Benni makes is that regressive assimilations resulting in a voiced sequence are rare, e.g., liczba [ljidba] ‘number’,61 when compared to assimilations to a voiceless sequence, e.g., babka [bapka] ‘grandmother’, grubszy [grup ] ‘thicker’, łowczy [wft ] ‘hunting, Adj.’, babski [bapsci] ‘female, derog.’, radca [rattsa] ‘counsellor’. This discrepancy, however, need not be significant from the phonological point of view for a number of reasons. Firstly, the smaller number of instances of assimilation to a voiced consonant is determined by the existing types of morphological affixes in Polish. The regressive assimilations almost always involve morphological concatenations62 – the second obstruent is added by a morphological process. Once the domain of assimilation is extended beyond the phonological word, that is, to external sandhi contexts, the assimilations to a voiced consonant are equally || 61 Recall the discussion under (3) in Chapter 1, and see the following footnote. 62 This cannot be said about the three forms in which both consonants belong to the root and they come together as a result of the vowel-zero alternation, for example, dech / tchu [dx ~ txu] ‘breath, nom.sg./gen.sg.’, wies / wsi [vj ~ fi] ‘village, nom./gen.’, wesz / wszy [v ~ f ] ‘louse, nom./gen.’. Interestingly, all of them involve the underlying sequence C+vC–v and not C– v +v C , and all of the clusters end in a fricative.

84 | Progressive voice assimilation in Polish

numerous because at this level they are determined by syntactic concatenation. The chances of finding a word beginning with a voiced obstruent are much higher than those of finding an affix beginning with a voiced obstruent. Secondly, the incidence of word-internal clusters in which we are dealing with static agreement does not seem to show such disproportions. Assimilations produce results which are parallel to the static cases of voice agreement of clusters of two obstruents, in e.g., kto [kt] ‘who’, and gdy [gd] ‘when’. Nevertheless, the rare occurrence of regressive assimilations to a voiced consonant as an observed process due to voice alternations within a root is a fact. It should also be noted that the cases of static voice agreement always involve consonants belonging to the same root, and in the absence of evidence in the form of alternations it is difficult to assume, other than for the economical reason of subsuming the static voice agreement and voice assimilation in one generalization, that the voice agreement is due to an active assimilation rule. The inclusion of Progressive Voice Assimilation in forms like twarz [tfa ] ‘face’, bitwa [bjitfa] ‘battle’, krzak [k ak] ‘bush’, and chrzan [x an] ‘horseradish’ in the discussion points to an implicit assumption that [ ] and [f] are devoiced // and /v/, respectively. The latter occur in voiced clusters such as drzewo [dv] ‘tree’ and dwa [dva] ‘two’. Thus, Benni assumes, without argument, that there is such a process as Progressive Voice Assimilation and that the targets are non-sonorant in nature. A schematic presentation of the word-internal assimilations is given below (after Benni 1909: 23). Note that in the case of regressive assimilation ([←]) the second consonant is provided by morphology, while in progressive assimilation ([→]), the two consonants are part of the same root. (4)

Consonant sequence a. C+v + C–v b. C

–v

+ C

+v

Result of assimilation =

Type of assim.

Example

C–vC–v

[←]

babka [bapka] ‘grandmother’

–v –v

[→] [←]

kwiat [kfjat] ‘flower’ prośba [pr ba] ‘request’

= i) C C = ii) C+vC+v

Benni notes that the existence of assimilation in both directions in (4b) could not be the result of voice specifications of the consonants and that it follows, rather, from other properties. He singles out occlusion and friction as the relevant properties and extends (4) to include the variation based on manner. Following Benni (1909: 23), the symbols T and S will be used below in (5) to define stops and fricatives, respectively.

Introduction | 85

(5)

Consonant sequence a. S+v + S–v

Result of assimilation

Type of assim.

Example

=

C–vC–v

[←]

wchód [fxut] ‘entrance’

=

–v –v

[←]

wtedy [ftd] ‘then’

–v –v

+v

b. S + T

–v

+v

–v

=

C C

[←]

babka [bapka] ‘grandmother’

d. T+v + S–v

=

C–vC–v

[←]

grubszy [grup ] ‘thicker’

+v

=

–v –v

C C

[→]

świnia [fjia] ‘pig’

f. S–v + T+v

=

C+vC+v

[←]

prośba [pr ba] ‘request’

g. T–v + T+v

=

C+vC+v

[←]

jakby [jagb] ‘as if’

=

–v –v

[→]

kwiat [kfjat] ‘flower’

c. T + T

–v

e. S + S

–v

h. T + S

+v

C C

C C

Note that the only two cases of a voiced result involve regressive assimilation (5f,g), and that progressive assimilation always results in a voiceless cluster. Benni proposes to deal with the above pattern by proposing the following two generalizations (p. 24). (6) The second consonant determines the voice quality of the cluster a. if it is occluded – it wins b. if it is a fricative – the whole group is voiceless The generalizations are possible mainly because of the assumed distinction between word-internal phonology and external sandhi, in which, however, voiced fricatives do influence the voicing value of preceding stops as well as of fricatives, e.g., brat wiedział [brad vjd aw] ‘brother knew’ and nasz wujek [na vujk] ‘our uncle’. However, we are still left with a number of questions. Firstly, why does progressive assimilation involve only /v/ and //? Secondly, can we safely assume that word phonology is markedly different from that in external sandhi in which the two fricatives impose voicing on the preceding consonant, e.g. brat wiedział [brad vjd aw] ‘brother knew’ and brat rzucił [brad utiw] ‘brother threw’? And finally, are we really dealing with assimilation in the case of words like świnia [fjia] ‘pig’, kwiat [kfjat] ‘flower’, krzak [k ak] ‘bush’, and trzeba [t ba] ‘one should’? More importantly, the generalizations in (6) seem to omit T+v + S+v, which are voiced, e.g., grzyb [g ] ‘muschroom’, and dwa [dva] ‘two’ and contradict (6b).

86 | Progressive voice assimilation in Polish

1.2 Is PVA dead or alive? There are a number of reasons that contribute to the conviction of linguists that progressive voice assimilation must be viewed as a living process in Polish.63 Their totality is only mildly convincing and proposals rejecting the living status of this phenomenon have also been put forward. Benni (1932) concludes that the two types of assimilations cannot really be viewed as one process. He distinguishes between living and real assimilations where two consonants of opposite voice value become adjacent, and dead assimilations which are no longer active. The living assimilations are the regressive ones and include babka [bapka] ‘grandmother’ and liczba [ljidba] ‘number’, while the dead ones are found in, for example, twarz [tfa ] ‘face’ and relate to the so called progressive assimilations. Benni concludes that it is probably the spelling difference – the use of ‘w’ and ‘rz’ in twarz, krzak, etc. – that had lead himself and other linguists astray. Indeed, when considering the possible arguments in favour of the position that progressive assimilation lives, spelling does not constitute particularly compelling evidence though it may have influenced the assumption that many linguists have made. A much stronger argument is based on living alternations, such as those below. (7) a. marchew [marxf] ‘carrot’

marchwi [marxfji] ‘gen.’

b. cerkiew cerkwi [tsrcf] [tsrkfji] ‘Orthodox ‘gen.’ church’

marchewka [marxfka] ‘dim.’

vs. marchewek [marxvk] ‘dim.gen.’

cerkiewka [tsrcfka] ‘dim.’

vs. cerkiewny [tsrcvn] ‘adj.’

[f] appears word-finally, before a voiceless stop and after a voiceless consonant, while [v] only appears intervocalically or before a sonorant. Benni downplays the importance of the fact that the voiced [v] in marchewek and cerkiewny finds no explanation under the assumption that progressive devoicing is a dead phenomenon. It should be noted that, excluding the problematic case of putative progressive assimilation in marchwi, all the remaining contexts in which [f] or [v] occur in (7) are consistent with the pro-

|| 63 Whether it is a phonological or a phonetic process is another question, which will be taken up in our analysis at the end of this chapter.

Introduction | 87

posal that we are dealing with an underlying /v/. Firstly, it surfaces as [v] in the intervocalic context and before a sonorant, e.g., cerkiewny. Secondly, like other obstruents, it may be assumed to be devoiced word-finally in marchew, and assimilated to the voiceless obstruent in marchewka. In other words, the entire paradigm is parallel to, for example, that of baba [baba] ‘woman, derog.’, bab [bap] ‘woman, derog.gen.pl.’, babka [bapka] ‘grandmother’. Thus, what would seem to remain unexplained is just the progressively devoiced [f] in marchwi. We are back to square one. The distinction into dead and live assimilations is criticized by Nitsch (1932) who admits that, possibly, the assimilation is dead in those cases where ‘w’ and ‘rz’ always directly follow the preceding consonant, e.g., twardy [tfard] ‘hard’ and krzywy [k v] ‘uneven’. Nitsch points to a difference between such forms and those in which the two consonants may be separated by an alternating vowel, such as in the forms in (7) above, and adds some more data (p.137). (8)

krew [krf] ‘blood’

krwi [krfji] ‘gen.’

krewny [krvn] ‘cousin’

brukiew [brucf] ‘swede’

brukwi [brukfji] ‘gen.’

brukiewek [brucvk] ‘dim.gen.pl’

krokiew [krcf] ‘rafter’

krokwi [krkfji] ‘gen.’

krokiewek [krcvk] ‘dim.gen.pl.’

rzodkiew [tcf] ‘radish’

rzodkwi [tkfji] ‘gen.’

rzodkiewek [tcvk] ‘dim.gen.pl.’

Some of the forms can exhibit the voiced [v] also in other types of morphological processes, for example, in augmentation, e.g., rzodkiewa [tcva] ‘big radish’, marchewa [marxva] ‘big carrot’. Such data, according to Nitsch, point to a psychologically voiced /v/ in the above forms and, consequently, to the existence of progressive devoicing. One should also add the aforementioned existence of non-assimilated forms, e.g., twarz [tva ] ‘face’ which occur in Kashubia, Great Poland and parts of Silesia as evidence that in the dialects which have [tfa ], the form is due to a process of progressive devoicing.

88 | Progressive voice assimilation in Polish

It is not possible to find similar representative data such as those in (7) and (8) to argue for the psychological // with respect to ‘rz’ in, e.g., trzeba [t ba] ‘one should’. The existing alternations are indirect, in that [ ] or [] alternate morphophonologically with [r], a voiced consonant, but also a sonorant. This brings us back to the unresolved distinction between obstruents and sonorants with respect to voicing. Nitsch (1932: 138) illustrates the relationships [r]–[]–[ ] with the following data. (9) a. góra [gura] ‘mountain’

górzysty [gust] ‘mountainous’

b. wydra [vdra] ‘otter’

wydrze [vd] ‘otter, dat.’

c. futro [futr] ‘fur coat’

futrze [fut ] ‘fur coat, dat.’

(9a) shows the regular morphophonological alternation between [r] and []. (9b) and (9c) demonstrate what happens with this alternation in the contexts of preceding voiced and voiceless consonant respectively. What connects these forms is obstruentization of /r/, and what distinguishes them is the presence or absence of voice assimilation. Nitsch concludes that a uniform phonetic principle should be assumed here, and suggests that proper examples should be used to illustrate Progressive Voice Assimilation, such as, e.g., krwi, krokwie, chytrze, ostrze, and not twarz, swój, kwiat, trzaska, in which no alternation is observed. It is not clear if the ‘false’ examples should therefore not be treated as assimilations in his view. Finally, he mentions in passing that the absence of a uniform principle for all types of assimilation (progressive and regressive) can be explained by diachronics, but he provides no clues. It is evident that it is the historical development of Polish that has singled out the two fricatives // and /v/ as targets of PVA. It also seems possible that some explanation for the progressive rather than regressive direction can be sought in the nature of obstruentization. The answer will depend on the relation between this process and the representation of voicing, as well as on the representational distinction between the /v/’s occurring in the data above and other /v/’s which are found in modern Polish, for example, word-initially. The last point relates to the fact that one has to explain the different behaviour of the /v/’s in, e.g., cerk[fj]i – cerkie[f] – cerkie[v]ek, in which the /v/ assumes the voice value of the preceding obstruent in cerkwi, and the /v/ which imposes its voice value on the preceding obstruent in brak wody [brag vd] ‘absence of water’. A valid question is whether we are dealing with different /v/’s in these positions, or perhaps with different /k/’s, or perhaps with something else again. In what follows, a functional explanation of progressive voice assimilation is provided.

Introduction | 89

1.3 A functional explanation of progressive devoicing Stieber (1946a,b) provides two brief historical sketches of the development of Tv/Tvj and Tř in Polish and offers a structural explanation of why these sequences favoured progressive voice assimilation instead of the expected and regular regressive one. There are two respects in which the shifts Tv → Tf and Tř → T differ. Firstly, they occurred at different time in the history of Polish. Stieber assumes that the former took place sometime in the 12th century while the latter occurred sometime at the turn of 17th-18th century. The second difference concerns the geographical scope of the two phenomena. While Tř → T occurred in practically all dialects of Polish, Tv → Tf is limited to Mazovia, Lesser Poland, South-East Upper Silesia and the former Łęczyca and Sieradz regions. The voiced fricative [v] in Tv contexts, e.g., [tvja], is retained in Kashubia, Great Poland and the remaining part of Silesia (Dejna 1973: 100). Despite the temporal and geographical differences mentioned above, Stieber (1946a,b) manages to provide a unified structural explanation for the Progressive Voice Assimilation, where it did occur, which is as follows. At the time of the respective shifts Tv → Tf and Tř → T , there was no phonemic distinction / or v/f in this particular context in Polish, and, therefore, the effect of progressive assimilation to the voiceless fricative, that is, → and v → f was less troublesome functionally speaking – for example, it produced no homophony – as opposed to the potential consequences of the expected regressive voice assimilation, in which the voiced fricative would impose its voice value on the preceding voiceless consonant (10). The data below are taken from Stieber (1946a,b), and show some potential homophonic pairs between the reflexes of Tř and Dř, as well as Tv and Dv, if regressive voice assimilation had taken place in the T_ context. (10)

Potential homophonic pairs a.

Dř

D  *D

grzywa [gva] ‘mane’, grze [g] ‘game, dat.’ drze [d] ‘he/she tears’ vs.

Tř T

krzywa [k va] ‘uneven, fem.’ krze [k ] ‘iceberg, dat.’ trze [t ] ‘’he/she rubs’

90 | Progressive voice assimilation in Polish

b.

Dv

Dv *Dv

dwoje [dvj] ‘two people’ dwór [dvur] ‘court’ drwale [drval] ‘woodcutter, pl.’64 vs.

Tv Tf

twoje [tfj] ‘your’ twór [tfur] ‘creature’ trwale [trfale] ‘permanently’

Thus, it seems that Stieber assumes that first there was obstruentization to  and v, respectively, which produced sequences of two obstruents disagreeing in voicing. These sequences, then, were eligible input to regressive voice assimilation, but the structural considerations reversed the direction of assimilation to a progressive one. The uniformity of the functional analysis, despite the fact that the phenomena did not take place at the same time in the history of Polish, or in the same areas, is quite appealing from the theoretical point of view. However, a number of questions need to be addressed in connection with the fact that in some parts of Poland, the modern dialects still do not have either progressive devoicing or Regressive Voicing Assimilation in the case of Tv sequences, a situation which persisted as it was in the 12th century. This does not overthrow Stieber’s analysis because the absence of progressive assimilation still does not create homophony, while a possible regressive assimilation would. The potential regressive assimilation in twarz [tva ] ‘face’ would also make such vocabulary items sound very different from the other dialects, e.g., *[dva ].

2 Modern derivational and non-derivational accounts of PVA 2.1 Preliminaries The modern generative views on the so called progressive voice assimilation (PVA) are based on a variety of theoretical assumptions to do with the basic concepts such as obstruent and obstruentization, phonological representation

|| 64 The pair drwale / trwale is interesting as it contains an intervening sonorant and bears on the issue of sonorant transparency in Polish, to be discussed in Chapter 5.

Modern derivational and non-derivational accounts of PVA | 91

of the voicing contrast, possible and impossible phonological processes (rules) and their organization in grammar, as well as the very nature of the interaction between phonology and phonetics on the one hand, and phonology and morphology on the other. It should be added that the very name of the phenomenon is slightly misleading given that authors sometimes refer to it as progressive devoicing (PD), and avoid postulating a rule of progressive spreading of the voicing property (Gussmann 1992; Rubach 1996), leaving the last stage of the derivation to default filling rules.65 Thus, the most basic question to do with this phenomenon should be whether we are dealing with a phonological process at all. Nevertheless, regardless of the theoretical perspective taken, some intuitions and assumptions persist as they revolve around the basic facts to do with the Progressive Voice Assimilation. One such intuition is that the fricatives /v/ and // exhibit something which might be referred to as sonorant-like behaviour (Bethin 1992; Gussmann 1992; Rubach 1996). There are two or three reasons for such an assumption. Firstly, there is historical evidence, which was evoked above, that the labio-dental fricative in the relevant contexts comes from a labio-velar glide /w/, while the post-alveolar fricative // evolved from a palatalized /r/. Secondly, the modern Polish alternations between /r/ and //, in, e.g., góry / górzysty [gur ~ gust] ‘mountains / mountainous’, suggest an affinity of the fricative // to the sonorant /r/. Although parallel alternations involving the labio-dental fricative [v] and labio-velar glide [w] are not on record, the existence of dialectal forms such as marchwi [marxvji] ‘carrot, gen.’, in which the fricative remains voiced, is also strong evidence for the sonorant-like behaviour of the fricative. It should be noted that implicit in the phrase ‘sonorant-like behaviour of the fricative’ is the assumption that we are dealing with an obstruent (fricative = obstruent). Recall that two obstruents must agree in voicing in Polish. Thus, forms like twój [tvuj] ‘your’ and marchwi [marxvji] ‘carrot, gen.’ violate that phonotactic regularity unless they are indeed not cases of two obstruents.66 Thus, the [v] is not devoiced, and does not voice the preceding obstruent because it ‘behaves like a sonorant’, whatever that means. Interestingly, the same fricative exhibits regular obstruent-like behaviour word-finally, where it is devoiced (marchew [marxf] ‘carrot, nom.’) and before affixes beginning with a voiceless obstruent (marchewka [marxfka] ‘carrot, dim.nom.sg.’). In this light, the [v] in marchewek

|| 65 Bethin (1992: 176), following Mascaró (1987), assumes spreading which takes place before obstruentization (cf., laryngeal control in Gussmann 2007). 66 Alternatively, one should perhaps reconsider the existence of active voice assimilation in obstruent clusters inside roots in Polish.

92 | Progressive voice assimilation in Polish

[marxvk] ‘carrot, gen.pl.’ is ambiguous. Is it still a sonorant or is it already an obstruent, just not devoiced by context? Its voiced nature is consistent with both assumptions. It transpires that a clear theoretical view is required concerning the concept of obstruency in phonological theory and the nature of obstruentization with relation to the voicing phenomena. The relevant questions are listed below: (11)

Obstruents and obstruentization a. does a non-obstruent (sonorant) become an obstruent with respect to the voicing phenomena at hand, and if so, at what stage? b. how is the change effected in the representation? c. is there a phonological category, say, [±obstruent], or does something else decide the status of  and v? d. is ‘obstruent’ a phonological or perhaps a phonetic property/category?

Connected with the status of an obstruent and its laryngeal specification is the typical understanding of the Progressive Voice Assimilation as primarily the loss of laryngeal specification on the fricatives /v/ and // in post-obstruent context.67 Interestingly, this analysis does not really require a reference to a sonorant-like behaviour of the fricatives. Some details of the relevant proposals are given below. Finally, very much like Benni (1932), who claimed that PVA is a dead phenomenon, a number of generative analysts, one way or another, express the intuition that progressive voice assimilation is not an active process. In the review of the most important contributions we will pay attention to the four assumptions discussed above, that is, i) sonorant-like behaviour of the fricatives, ii) the formulation of the process as delaryngealization of the fricative, iii) the status of PVA as a live phonological process, and iv) the difference between PVA in Tv contexts and RVA (Regressive Voice Assimilation) across word boundaries (...T#v...). As in our earlier discussion of the representation of voicing and assimilation processes in Polish, we will pay attention to such theoretical points as privative versus binary representation of voice, the formulation of the processes and the phonetics – phonology relation.

|| 67 This analysis is also used in, for example, Iverson and Salmons (2003b) who propose a rule of ‘post-obstruent fricative neutralization’ to deal with a very similar phenomenon in Dutch.

Modern derivational and non-derivational accounts of PVA | 93

2.2 Gussmann (1992) In Gussmann (1992) all voicing phenomena in Polish are directly derived from syllabification and processes affecting syllable structure, especially resyllabification, as well as principles governing the licensing of voice autosegments by the syllable node – the Voice Licensing Principle. Gussmann assumes that at the phonological level (phonological representation) the maximal syllable structure in Polish observes the CRVR template, which itself follows from the Sonority Sequencing Generalization. It contains a branching onset, a vowel and a sonorant coda. Obstruents can occur only in onsets, while sonorants can occur in both non-syllabic constituents, that is, onsets and codas. Obstruents are marked for voice contrastively at the lexical level and carry either [+voice] or [– voice]. Sonorants are not specified for voice phonologically, yet they receive a non-contrastive specification [+voice] quite early by phonological derivation.68 It is important to add that only syllabified sonorants have their [+voice] licensed by the syllable node. The syllable node licenses only one occurrence of the contrastive [voice] autosegment. This, coupled with the restriction that obstruents are never syllabified in the coda, produces the desired result that obstruents must agree in voicing. This is because non-syllabified obstruents eventually end up in the so called complex onset if they are word-internal, or remain unsyllabified if they are word-final. The unsyllabified final consonants lose their underlying voicing specification and later in the derivation they receive [–voice] by Default Filling. Let us clarify the concept of the complex onsets. They involve two or even three obstruents and are not underlying. They are always derived by resyllabification rules which incorporate unsyllabified obstruents into existing structures.69 The unsyllabified obstruents in Gussmann’s model result from, e.g., vowel deletion rules, depriving underlying onsets of their nucleus. Before we look at an illustration of the workings of the model let us list the main mechanisms that Gussmann (1992) refers to in his analysis of voicing phenomena in Polish. They are given in the order in which they should apply. Some of the mechanisms are additionally commented on.

|| 68 The non-contrastive use of [+voice] is theoretically problematic. Distinctive features by their nature, if present, are meant to provide contrasts. 69 The complex onsets in Gussmann’s model should not be confused with branching onsets, as in, e.g., trawa [trava] ‘grass’, which are also sometimes referred to as complex onsets in the literature. To avoid this kind of confusion, we will use the term double onset to refer to the complex (derived) onsets.

94 | Progressive voice assimilation in Polish

(12) Polish Voice Licensing (PVL) (p. 43) The syllable node [σ] licenses (a) [±voice] on syllabified obstruents and (b) [+voice] on syllabified sonorants.

Sonorant Voicing Underlyingly unspecified sonorants, quite early in the derivation, are supplied with the specification [+voice] if they are syllabified.

Vowel Deletion Empty vowels are deleted if not followed by another empty V in the following syllable. This rule desyllabifies the consonant(s) in the underlying onset.

Voice delinking Automatic result of losing a syllabic affiliation (as per PVL).

Resyllabification A process of incorporation of unsyllabified consonants into onsets. In fact, only obstruents are resyllabified this way, leading to the creation of double onsets.

Obstruent-to-Obstruent Spreading (p. 44) [±voice] | [+obstr] [+obstr] This rule must crucially apply before Default Filling and is restricted to word-internal context.

Obstruent-to-Consonant Spreading (p. 53)

[+cons]

[±voice] | [+obstr]

This is an optional rule to account for voicing assimilations across word boundaries. It affects obstruents as well as sonorants, e.g., wiatr [vjatr] ‘wind’ vs. wiatr zachodni [vjadr zaxdi] ‘western wind’. This rule is extended in Cracow-Poznań Polish (CP) to ‘Voiced Segment-to-Consonant Spreading’ to account for the CP sandhi voicing discussed in Chapter 3.

Modern derivational and non-derivational accounts of PVA | 95

Obstruentization /w/ →/v/, /rj/ →// Default Filling (p. 43) A consonant which is unspecified for voicing receives [–voice]. This concerns both obstruents and sonorants.

Let us illustrate the above with an example derivation of tchu [txu] ‘breath, gen.sg.’ which alternates with dech [dx] ‘breath, nom.sg.’, thus suggesting that the first obstruent is lexically voiced. The underlying voiced /d/ in tchu is followed by an empty V, which is deleted by rule if not followed by another empty V. When the empty V is present in the following syllable, then the first V is realized as [], a phenomenon known as the Lower Rule (e.g., Gussmann 1980). Once the empty V is deleted, the preceding obstruent ceases to be an onset and its voice autosegment must be delinked too.70 (13)

a.

b.

σ O | C | T | +v

c.

σ R | V

O | C | X | –v

σ R | V | u

Ÿ

| C | T  +v

O | C | X | –v

σ R

O R | V | u

Ÿ

O | C | T

O | C | X | –v

V | u

The delinking which is observed in the second stage (13b) is due to the Voice Licensing Principle which licenses the voice autosegment only in syllabified consonants. Finally, a rule of Obstruent-to-Obstruent Spread applies, which

|| 70 Interestingly, this rule sequence does not seem to, and in fact cannot, work in words like kra [kra] ‘ice floe’ (cf., kier [cr] ‘ice floe, gen.pl.’ showing that an empty V is also present in such forms) and especially in gra [gra] ‘game’ (cf., gier [r] ‘game, gen.pl.’). In both cases the V is deleted but instead of voice delinking the sequence Obstruent + Sonorant is assumed to be integrated into one branching onset with the underlying specification for voice intact. The reasons for this are not discussed in Gussmann (1992).

96 | Progressive voice assimilation in Polish

operates within the new double onsets. Final obstruent devoicing is similar in all details except that the [–voice] is assigned by Default Filling rather than by spreading. Returning now to the problem of Progressive Voice Assimilation, the potential bleeding relationship between PVA and RVA in the cluster of obstruents is avoided by resorting to representing the relevant fricatives as sonorants /rj/ and /w/ respectively. The sonorants are turned into obstruents at the relevant moment in phonological derivation by context-free rules /rj/ → // and /w/ → /v/ respectively. A slightly adapted derivation of krze [k ] ‘ice floe, dat.’ is illustrated below. (14a) illustrates the stage after [+voice] licensing in the sonorant but before empty vowel deletion and palatalization of /r/. The latter two processes produce stage (14b). Later, the context-free rule turns /rj/ into //.71 The new obstruent sequence has to be resyllabified as a double onset. At this point, two contrastive voice autosegments cannot be licensed and for reasons which are not entirely clear it is the new obstruent that delinks its [+voice] (14d). The fricative later receives [–voice] by Default Filling. Gussmann does not propose a rule of progressive spreading. On the other hand, in the derivation of grze [g] ‘game, dat.’ the obstruentization does not produce two distinctive voice autosegments, so the two occurrences of [+voice] are said to conflate. (14) a.

b.

c.

d. σ

σ

σ

O R O R | | | | C V C V | | | K r  | | +v –v

σ O

Ÿ

R | C C V | | | K rj  | | –v +v

O

Ÿ

O | C | K | –v

σ R

O | C V | | Š  | +v

O

Ÿ

O | C | K | –v

R

O | C V | | Š   +v

Thus, neither progressive voicing nor devoicing is technically required. PVA crucially depends on obstruentization and the resyllabification that this change

|| 71 // in Gussmann’s derivations is represented as the archiphoneme /Š/ with [+voice].

Modern derivational and non-derivational accounts of PVA | 97

in status effects (sonorant → obstruent). PVA occurs in the case of ‘new’ obstruents, while RVA takes effect in the case of two original obstruents as in tchu. The derivation of bitwa [bjitfa] ‘battle’ with the labio-dental fricative is exactly parallel to that of krze. On the other hand, for dialects in which this word is pronounced with a voiced fricative, i.e., [bjitva], the obstruentization rule /w/ → /v/ is assumed to apply in the normal fashion in rhymes. However, in onsets it is assumed to apply late, at a stage when the Voice Licensing Principles no longer hold. The difference in the behaviour of word phonology, e.g., krze [k ] ‘ice floe, dat.’, bitwa [bjitfa] ‘battle’ on the one hand, and phrasal phonology, e.g., nurt rzeki [nurd ci] ‘river current’, brat widział [brad vjid aw] ‘brother saw’ in which assimilation is regressive is not referred to directly. However, it is rather clear how this distinction would be derived. Simply, word-final obstruents remain unsyllabified in Gussmann’s model.72 Thus, in the case of brat widział the final /t/ in brat cannot form the double onset with the following fricative (new obstruent), which was the condition under which the fricative would lose its [+voice] specification. Finally, let us look at the derivation of krwi [krfji] ‘blood, gen.’, which alternates with krew [krf] ‘blood, nom.’ and krewny [krvn] ‘relative’. These forms tell us two things. Firstly, the fricative is devoiced in krew, and the branching onset /kr/ is underlyingly followed by an empty vowel which has to be deleted in krwi, with all the consequences of the move. The diagrams are adapted from Gussmann (1992: 51). (15) a.

b.

c. σ

σ

σ

R O R | | | C V C V | | | r w i | | +v +v

O

R

Ÿ

O | C C C V | | | | k r w i | –v +v

O

O

O C | K | –v

σ R

O

Ÿ

O | C C C V | | | | K r F i | –v +v Ÿ ∅

|| 72 Gussmann (1992: 40) views this as an argument against the principle of Stray Erasure.

98 | Progressive voice assimilation in Polish

The analysis is as follows. The representation in (15a) still contains the empty V which is responsible for the vowel-zero alternation in krew / krwi. The two sonorants already contain [+voice] as per Sonorant Voicing. Upon the empty vowel deletion (15b), the word becomes monosyllabic and contains a double onset involving two specifications for voice: [–voice] in the obstruent and a conflated [+voice] in the two sonorants. When the rule of obstruentization /w/ → /v/ turns the glide into a fricative it must lose its specification, being the ‘new’ obstruent in a double onset. Because the feature is shared by two consonants, both of them lose it. Later in the derivation they receive [–voice] by Default Filling. There is a fundamental problem with this analysis, which was noted already in Rubach (1996).73 It is not clear, as in the case of krze, why vowel deletion does not produce the automatic delinking of the voice specifications in the onset /kr/, especially since this is the course of action that Gussmann proposes in the analysis of a parallel structure of płci [pwti] ‘gender, gen.’ and mędrka [mntrka] ‘wiseacre, gen.’ when accounting for word-internal sonorant devoicing and obstruent voice assimilation across such sonorants (cf., mędrek [mndrk] ‘wiseacre, nom.’). Thus, technically speaking, the representation in (15b) should look as shown in (16) below. One should expect that the resyllabification of this configuration into a double onset leaves the first two consonants unspecified for voice. The obstruentization /w/ → /v/ would then feed Obstruent-to-Obstruent spreading, thus yielding the incorrect form *[grvi], or in fact *[grvi] because the resyllabification and the voice spreading involve only obstruents. The intervening sonorant should remain unsyllabified and receive [–voice] by Default Filling. (16)

σ

| C | K  –v

| C | r  +v

O | C | w | +v

R | V | i

|| 73 An extensive analysis of this and other problems with Gussmann’s analysis can be found in Rubach (1996).

Modern derivational and non-derivational accounts of PVA | 99

To salvage the analysis one might propose that the sequence /Kr/ in (16) receives [–voice] by Default Filling before the construction of the double onset, in which case further derivation would proceed as in (15) above with the minor exception that the sonorant /r/ would already be devoiced and PVA would involve delinking of [+voice] under the ‘new’ fricative (for the same reasons as Gussmann proposed) and specifying it with [–voice] by Default Filling again. The obvious problem with this amendment is that the Default Filling would have to be applied twice – once before double onset formation and the second time after the delinking of voice on the ‘new’ fricative – only to salvage the specification in the first obstruent. There is another reason why this move would not be beneficial as is shown by forms like brwi [brvji] ‘eyebrow, pl.’ and drwal [drval] ‘lumberjack’ corresponding to brew [brf] ‘eyebrow, sg.’ and a related form drewno [drvn] ‘timber’. If the vowel deletion rule desyllabified the first branching onset leading to the delinking of voice specifications and Default Filling indeed provided voice specification before double onset formation, then the sequence /Dr/ would have [–voice] and would likewise enforce [+voice] delinking on the ‘new’ fricative, which would later receive [–voice] by Default Filling. The result of these derivations would be *[prfi] and *[trfal]. Thus, some other mechanism is necessary to avoid voice delinking in onsets followed by sonorants. Only this course of action would salvage the analysis based on sonorant obstruentization, resyllabification and voice licensing.

2.3 Bethin (1992) Bethin (1992) is in many respects similar to the proposal in Gussmann (1992), although it seems she manages to bypass some of the problems of the latter. The voicing phenomena in Polish are syllable governed, with the template CRVR playing a crucial role. Secondly, only obstruents in onsets retain voice specification. The fricatives in question are phonological sonorants which are obstruentized in the course of phonological derivation. They end up in a compound onset, which is a similar construct to the complex (double) onset in Gussmann’s analysis. The new fricatives lose their voice specification because they have changed their major class feature and consequently their syllabic affiliation.74 The analysis is best summarized in Bethin’s own words:

|| 74 Cf., the concept of ‘new’ obstruents in Gussmann (1992).

100 | Progressive voice assimilation in Polish

Progressive assimilation may be seen as a consequence of syllabification principles. It involves the following steps: 1) obstruentization, 2) delinking from syllable structure, 3) loss of voice, and 4) attachment to onset obstruent with the concomitant sharing of voice specification. (Bethin 1992: 177)

The syllabification is slightly different, though. For example, there is no initial syllabification of onsets based on yer vowels (the empty V), and so there is no question of delaryngealization upon the deletion of the empty vowel. Additionally, word-final obstruents are eventually syllabified in the rhyme. Another important difference lies in the featural representation of voice. Unlike Gussmann, Bethin uses a privative system of the representation of the voice contrast, although the privative distinctions hold only at early levels of derivation and privativity does not affect the analysis of PVA. Bethin follows Mascaró (1987) in assuming that in a privative system like Polish there are three stages of voice specification. At Stage 1 only the [+voice] feature is present in obstruents. The voiceless obstruents are supplied with the feature [–voice] at Stage 2, while sonorants receive [+voice] at Stage 3. Thus, sonorants, albeit late, are still marked as [+voice] phonologically and may participate in assimilation phenomena, but only at the relevant stages in the derivation. Mascaró’s stages of voice specification are illustrated below on the basis of Polish /b, p, w/. We use symbols C and S for obstruents and sonorants respectively. The specifications are in the superscript. Non-specified objects have a superscripted zero (‘o’). (17) /b/ /p/ /w/

Stage 1

Stage 2

Stage 3

C+voi Co So

C+voi C–voi So

C+voi C–voi S+voi

Progressive Voice Assimilation in Polish takes the form of spreading [–voice] from the obstruent to the following sonorant at Stage 2, when the voiceless obstruents already contain the relevant feature, but the sonorants are still unspecified.75 Thus, Mascaró also assumes a lexical sonorant in the case of bitwa and krze. The sonorants are voiceless at the point of the operation of the spellout rule which obstruentizes /rj/ and /w/. Since the desonorantization merely

|| 75 The rule of [–voice] spreading is general and is assumed to apply equally to forms like bitwa ‘battle’, which are of interest here, and to other instances of voiceless obstruents followed by sonorants (C–voiSo → C–voiS–voi), e.g., jasny [jasn] ‘bright’, except that the effects of the spreading are later to be undone by Default ([+son]→[+voi]).

Modern derivational and non-derivational accounts of PVA | 101

changes the feature [+son] to [–son] and does not manipulate voicing, the progressive assimilation is already in effect and no additional progressive obstruent delinking and spreading is necessary. A schematized derivation of bitwa and krze is given below.76 /bj i t–voi

(18) Spreading

j

/b i C

–voi

Desonorantization /bj i C–voi j

Phonetic

[b i t

wo

a/

/k–voi

–voi

a/

/C

–voi

C–voi

a/

f

a]

S

rjo

/

–voi

/

/C–voi

C–voi

/

[k

]

S

Bethin modifies Mascaró’s analysis. For her, progressive spreading is also involved, and it is also placed at Stage 2 of the specification for voice. However, before the spreading takes place, the sonorants are first obstruentized and subsequently delaryngealized. Only then does spreading of both values take place in, e.g., krze [k ] ‘ice floe, dat.’ and grze [g] ‘game, dat.’, respectively. Bethin proposes that both [k ] and [g] are results of progressive assimilation. It will be recalled that in Gussmann (1992), progressive devoicing involves voice deletion and then default assignment of [–voice], while forms like [g] involve conflation of [+voice]. Thus, it looks like Bethin does not require reference to the sonorant nature of the target fricatives, other than to effect their delaryngealization upon obstruentization as per the Voice condition in order to render them targets to voice spreading. In sum, Bethin uses the following major mechanisms to derive not only PVA but also RVA. (19)

Voice condition (p. 167) Onsets preserve their specification for voicing. N” N X | root [–son] [+cons] | [voice]

|| 76 The starting phonological form already contains a palatalized /rj/.

102 | Progressive voice assimilation in Polish

This condition disallows voice specification in codas and at the end of the word. It also deprives the obstruentized fricatives of their voice specification rendering them liable to progressive spread. (20)

Progressive Voice N” | X | root [+cons] [–son]

X | root [+cons] [–son]

[voice] This spreading takes place within compound onsets, which is an interesting effect because it gives Bethin an independent solution – based on syllabification and obstruentization (/w/ → /v/, /rj/ → //) – to the problem of potential bleeding relation between PVA and RVA, where the latter is taken care of by a general spreading rule. (21)

Voice Spread (Regressive Voice Assimilation) X1 | root [–son] [+cons]

X2 | root [–son] [+cons]

[voice] The two rules are in fact complementary because Progressive Voice is found in the configuration marked-unmarked, while regressive Voice Spread occurs in unmarked-marked sequences. It should be stressed that the two mirror image configurations are due to general syllabification rules and conditions. The problem of the difference in the behaviour of word and phrase level phonology, e.g., krze [k ] ‘ice floe, dat.’, bitwa [bjitfa] ‘battle’ versus nurt rzeki [nurd ci] ‘river current’, brat widział [brad vjid aw] ‘brother saw’ receives a rather neat account. For progressive voice assimilation to take place, we must be dealing with obstruentization within a well-formed underlying onset leading to the delaryngealization of the new fricative and its adjunction to a compound

Modern derivational and non-derivational accounts of PVA | 103

onset. These conditions are not met across the word boundary. For example, the obstruentization /w/ → /v/ in widział does not cause its desyllabification. It simply remains in its onset. As for forms like krwi [krfji] ‘blood, gen.’, it seems that the analysis would be to a great extent similar to the one offered for the forms in which no intervening sonorant is present, e.g., bitwa [bjitfa] ‘battle’. The intervening sonorant is simply assumed not to be syllabified at the time the voicing rules apply. The weak point of the proposal lies in the fact that privativity is assumed but everything seems to work like in a binary system, e.g., progressive voice spreading must assume that [–voice] is already there. This point is criticized in Rubach (1996) whose views on privativity are much stricter than those of Bethin, even though he is arguing against it. We agree with Rubach’s views on privativity, for example that one cannot refer to [−voice] in a system which does not have it. Privativity viewed strictly, as is the case in, for example, Laryngeal Realism and Laryngeal Relativism, is much easier to criticize. It should be noted, however, that privativity in Bethin’s model plays a marginal role. Most voicing features are supplied by rules in very early stages of the derivation, including that on syllabified sonorants, and Bethin is very clear on this point, that both values will control the effects of progressive spreading.

2.4 Rubach (1996) From the above discussion it is rather clear that two main theoretical aspects, namely, syllable based generalizations and privativity, do not seem to work well in the voicing complex of Polish. In reaction to these analyses, Rubach assumes that voice assimilation in Polish is not connected to syllable structure but rather to linear adjacency of laryngeal nodes and specific rules defining the behaviour of such nodes.77 He also argues for a binary representation of voice. PVA is due to a Progressive Delinking (PD) rule, which delinks the laryngeal node in a fricative after voiceless consonants.

|| 77 Prosodic structure is crucial, however, in the analysis of transparency and opacity of sonorants, which are accounted for by referring to late and early adjunction respectively. Sonorants receive [+voice] by Sonorant Default which assigns this feature only to prosodified sonorants (cf., the discussion in Chapter 5).

104 | Progressive voice assimilation in Polish

(22)

Progressive Delinking R | L | [–voice]

R L

[–sonor]

[+cont]

The structural description of the PD rule clearly shows that the target fricative is an obstruent. Rubach is sceptical about the existence of a /w/ → /v/ obstruentization rule, although he can easily incorporate it in the analysis by ordering it before the PD rule, on a par with the /rj/ → // obstruentization, about which he is not sceptical. The PVA effect in words like bitwa and krze is achieved not by an additional rule of progressive [–voice] spreading but by filling in the laryngeal specification on the delaryngealized fricative by a universal Voice Default. (23)

Voice Default R | [–sonor]

R →

L | [–voice]

[–sonor]

L = laryngeal node, R = root node It will be recalled that RVA is dealt with in Rubach (1996) by means of two rules: Obstruent Delinking, which, in a sequence of two obstruents, erases the laryngeal node in the left-hand obstruent (24a), and Spread, which spreads the laryngeal node to segments, in effect, to obstruents with a delinked node (24b). (24)

a.

b. Obstruent Delinking R

[–sonor]

Spread R

L

L

R [–sonor]

R L

[–sonor]

The potential bleeding relationship between PVA, which is derived by Progressive Delinking and Voice Default, and RVA, which is derived by Obstruent Delinking and Spread is resolved by placing the rule of Progressive Delinking in

Modern derivational and non-derivational accounts of PVA | 105

the postcyclic lexical component (word phonology) and the Obstruent Delinking in the postlexical component (phrase level phonology). In other words, the application of the two rules is ordered by placing them in two consecutive blocks of rules. Thus the rules do not even interact directly with each other. This distinction is also responsible for the difference between bitwa [bjitfa] ‘battle’ and brat Wandy [brad vand] ‘Wanda’s brother’. The rule of Progressive Delinking does not operate at the postlexical level. As for the difference between krze and grze, the former is derived by Progressive Delinking, while the latter does not fit the structural description of the rule so the fricative simply retains its [+voice]. Finally, krwi is derived in the same way as bitwa. This is possible due to the assumed ordering of prosodification of the intervening sonorant very late in the derivation – the last stages of the postlexical component. Then the sonorant will also receive [+voice] by Sonorant Default which supplies this feature on prosodified sonorants. It is this systemic voicing of sonorants that might be one of the drawbacks of Rubach’s otherwise comprehensive analysis. Rubach cannot derive voiceless sonorants in bóbr [bupr] ‘beaver’, bojaźń [bja] ‘fear’, Jędrka [jntrka] ‘Andrew, dim.gen.’, or krwi [krfji] ‘blood, gen.’, which is probably why this fact is never mentioned in the data. To be precise, it is not that this devoicing should be derived. Rubach himself admits (footnote 12, p. 80) that ‘the voicing or devoicing of sonorants of the type exhibited in (22) [forms like Jędrka] are best relegated to the realm of phonetic implementation rules.’ The problem is that Rubach’s mechanisms do not leave anything to phonetic interpretation. The sonorants are subject to late adjunction and then to Sonorant Default which provides an unambiguous and non-gradient feature [+voice]. The problem seems to be a little deeper still. It appears that the absence of sonorant devoicing in CrC clusters may sometimes bring about the absence of PVA. Thus, for example, speakers who pronounce krwi as [krfji] and Jędrka as [jntrka], which means that they follow the correct derivation with an unsyllabified transparent sonorant and Progressive Delinking, may still pronounce plwać as [plvat] and find [plfat] artificial.78 This shows the obvious connection between the phonetic shape of the sonorant and the following fricative.79 However, the interpretation of that connection is not so obvious. One

|| 78 In Karaś and Madejowa (1977: 302) the word plwocina ‘spit’ has two alternative pronunciations: [plftina] and [plvtina] as equal in status. A similar variation is found in Siedlce [tlts~ !lts] ‘town name’. 79 Steriade (1999), for example, attributes such effects to the relative duration of the sonorant. Whether the length of the sonorant is the cause of the variation or whether it is due to some

106 | Progressive voice assimilation in Polish

could propose that the phonetic shape of the fricative is dependent on the presence or absence of sonorant devoicing. This alternative proposal would be in fact more compatible with the fact that one speaker may have [krfji] and [plvat] at the same time. Gradience is indeed present in such forms, and it works against the derivational approaches, in which a fully specified systematic phonetic level is arrived at in the course of the derivation. Even though a lot of ink has been spilt in the literature to argue against the basic theoretical tenets used in this analysis, such as extrinsic rule ordering or binary representation of voice, with the exception of the problematic sonorant behaviour, it is difficult to point to weaknesses in Rubach (1996) when looking from within that system. Two, however, need to be mentioned. Firstly, the Progressive Delinking rule is accurate but rather arbitrary in its limitation with respect to the context and the target. The delinking occurs only after voiceless consonants. It is rather unclear why symmetry is not allowed here in a system with symmetrical voice assimilation of the regressive type (Obstruent Delinking), and interaction between laryngeal nodes, not features. The analysis also ignores, or avoids an explanation of the fact that the two fricatives which are targets in PVA are cases of obstruentization. Certainly, obstruentization is included in the derivation but it is tangential to the analysis of PVA. As we said above, the rule of PD is accurate in its application. However, the specification of the first consonant as voiceless avoids a wrong derivation in just a handful of existing instances of forms in which the first consonant is not voiceless. Incidentally, in these forms, the second fricative is not one of our obstruentized sonorants. The relevant cases are: dech / tchu [dx ~ txu] ‘breath, nom.sg./ gen.sg.’, wieś / wsi [vj ~ fi] ‘village, nom./gen.’, wesz / wszy [v ~ f ] ‘louse, nom./gen.’. Finally, Rubach (1996) does not mention or discuss the problem of unassimilated forms such as [tvuj], [marxvji], [krvji], which are not simple cases for analyses in which a fricative is automatically identified as an obstruent. One could, of course, suggest that the rule of Progressive Delinking is not present in the grammars of speakers who produce such forms, but these speakers would also have to lack the rule of Obstruent Delinking followed by Spread. This, however, is not an option. These speakers do apply the above rules to forms outside the PVA context.

|| other principle is not clear. See Chapter 5 for more discussion of this point. Although there seems to be an inherent length difference between, say, [r] and [l].

Modern derivational and non-derivational accounts of PVA | 107

2.5 Gussmann (2007) Gussmann (2007) is couched in a model similar to Laryngeal Realism (Harris 1994; Honeybone 2002). The model uses a privative representation in which the voice opposition in Polish is marked by the presence of the low tone element {L} in voiced obstruents and its absence in the voiceless unaspirated series. Additionally, the Element Theory used by Gussmann does not allow for laryngeal specification of sonorants at any stage of derivation. Gussmann aims to express the long standing intuition (e.g., Benni 1909) that PVA is part and parcel of phonotactic tendencies leading to what he calls voice uniformity in obstruent clusters in Polish. Paradoxically, and contrary to the main generalization about voice agreement, progressive voice assimilation – which concerns only two fricatives in Polish, namely, // and /v/ – operates inside words and does not work in juncture position, where these fricatives voice the preceding obstruent as expected, regressively, e.g., smak wina [smag vjina] ‘taste of wine’ and ptak rzeczny [ptag t n] ‘river bird’. The author points to two main reasons why PVA should be treated as an active phenomenon. The first one concerns the existence of alternations such as gra / grze [gra ~ g] ‘game, nom./dat.sg.’ and especially kra / krze [kra ~ k ] ‘ice floe / dat.sg.’. As for the alternation gra / grze [gra ~ g] ‘game, nom./ dat.sg.’, Gussmann proposes that the relation between [r] and [] is expressed by a morphophonological replacement. Thus it is a sonorant marked with the diacritic which ensures its replacement in the required contexts. (Gussmann 2007: 309)

He further adds in a footnote that this replacement is referred to as ‘obstruentization’ in derivational frameworks. Thus, the obstruentization /r/ → // in this analysis is not part of phonological derivation, but rather of morphophonological derivation which precedes phonology. An /r/ is replaced with an obstruent // in [g], and the direct consequence of this replacement is that the fricative (obstruent) must have its laryngeal specification, {L}. Since the sonorant is not replaced with / /, the voiceless fricative in krze [k ] ‘ice floe, dat.sg.’ must be a result of PVA.80 The // is subject to other regular voicing phenomena such as final devoicing, c.f., lekarza / lekarz / lekarski [lkaa ~ lka ~ lkarsci] ‘doctor, gen.sg./ nom.sg./ medical’. || 80 Note that the term PVA (Progressive Voice Assimilation) is only a term describing the fact that the fricative becomes voiceless in the relevant context T_. It does not refer to a particular process, but merely expresses the consequences of some process(es).

108 | Progressive voice assimilation in Polish

The second argument for assuming that PVA applies is based on the existence of the unassimilated forms such as twój [tvuj] ‘your’. Presumably, the logic of this argument is that the existence of [tv] in some varieties of standard Polish next to [tf] shows that in some varieties or idiolects the PVA phenomenon is not active, and therefore it is active in the varieties that have [tfuj]. This, however, concerns only the labio-dental fricative. One would immediately wish to ask what exactly happens in the unassimilated clusters and how PVA is effected in the assimilated ones, given the representation of the voicing contrast which Gussmann adopts. This brings us first to the problem of sonorant – obstruent distinction. Gussmann’s position on the sonorant-like behaviour of the fricatives is rather complex, as the following quote suggests. The alternations, and in particular those involving the alveolar fricative and the sonorant [r], lead us to propose an initial interpretation: the fricatives [ , ] are not fricatives at all but rather sonorants. (Gussmann 2007: 309)

The quote clearly shows that the term fricative is unambiguously equated with obstruents. This is particularly interesting in the light of the fact that Government Phonology does not recognize the labels ‘obstruent’ or ‘sonorant’. These are not phonological categories but rather impressionistic and convenient cover terms. On the contrary, ‘fricative’ can be related to phonological categories of at least three types. The first one is the so called ‘noise’ element {h} (Harris 1990). The second one is the presence of a high tone {H} element (Backley 2011; Cyran 1997: 191). The third way to express friction is to refer to a headed status of a resonance element (Cyran 1996; Ritter 1997). Thus, a phonetic labio-dental fricative in Polish may have two alternative representations: {U,h,L} and {U}. This is exactly what Gussmann has in mind. The simplex (mono-elemental) representation is a [v] which behaves like a sonorant.81 It does not affect the preceding obstruent regressively because it does not have a low tone to spread. Let us compare the representations of twój [tvuj] ‘your’ without regressive or progressive assimilation and smak wina [smag vjina] ‘taste of wine’ below. The exact syllabic representation is ignored, as it appears to be irrelevant.82 Only the fricative is given a full melodic representation.

|| 81 In fact, Gussmann does not mention headedness of the element {U} but merely its nonnuclear affiliation. 82 Since the relevant clusters will behave in the same fashion whether they form a branching onset or a sequence of onsets (Gussmann 2007: 310), we will use a CVCV skeleton here, as we do throughout this book.

Modern derivational and non-derivational accounts of PVA | 109

(25) a.

b. C V C V C V | | | | t U u j

C V C V C V | | | | s m a k | < < <