Phonetics and Phonology in Russian Unstressed Vowel Reduction

Phonetics and Phonology in Russian Unstressed Vowel Reduction: A Study in Hyperarticulation Abstract: Unstressed vowel reduction figures centrally in recent literature on the phonetics-phonology interface, in part owing to the possibility of a causal relationship between a phonetic process, duration-dependent undershoot, and the phonological neutralizations observed in systems of unstressed vocalism. Of particular interest in this light has been Russian, traditionally described as exhibiting two distinct phonological reduction patterns, differing both in degree and distribution. This study uses hyperarticulation to investigate the relationship between phonetic duration and reduction in Russian, concluding that these two reduction patterns differ not in degree, but in the level of representation at which they apply. These results are shown to have important consequences not just for theories of vowel reduction, but for other problems in the phonetics-phonology interface as well, incomplete neutralization in particular. Introduction Unstressed vowel reduction has been a subject of intense interest in recent debate concerning the nature of the phonetics-phonology interface. This is the case at least in part due to the existence of two seemingly analogous processes bearing this name, one typically called phonetic, and the other phonological. Phonological unstressed vowel reduction is a phenomenon whereby a given language's full vowel inventory can be realized only in lexically stressed syllables, while in unstressed syllables some number of neutralizations of contrast take place, with the result that only a subset of the inventory is realized on the surface. A representative example of such a pattern is found in many Eastern Bulgarian dialects (e.g., most of those classified as Central Balkan by Stojkov [1993: 108-111]). In these dialects, lexically stressed syllables realize a six vowel inventory /i, e, a, , o, u/, while lexically unstressed syllables are restricted to the three vowels /i, , u/.1 Phonological unstressed vowel reduction is thus one of a number of patterns of neutralization that Trubetzkoy [1969: 235-241] called “structurally conditioned”, and that Steriade [1995] has more recently termed “positional”. Positional neutralization patterns are characterized by the ability of certain structural positions within the word or phrase to license greater numbers of contrasts than other positions. What is usually called phonetic vowel reduction, on the other hand, is the phenomenon, investigated perhaps most famously by Lindblom [1963], whereby in certain contexts vowels fail to achieve what would otherwise be considered their acoustic targets, suffering instead some degree of phonetic “undershoot”. In his foundational work on the subject, Lindblom demonstrates a relationship between the phonetic duration allotted to the articulation of certain Swedish vowels and the degree of gestural undershoot to which those vowels are subject. Essentially, with less time available to carry out the necessary gestural routines, speakers fall short of target articulations, with the result, Lindblom argues, of greater coarticulation with neighboring consonants, and hence also formant patterns that fall short of the values typical of more favorable durational circumstances. Noting that unstressed vowels in phonological vowel reduction languages tend to be produced with curtailed durations relative to analogous stressed vowels, many phonologists have posited a connection of one sort or another between phonological vowel reduction on the one hand and phonetic reduction or gestural undershoot on the other [e.g., Barnes, 2002, 2006; Beckman, 1998; Crosswhite, 2001, 2003; Flemming, 1995/2002, 2001, 2005; Herrick, 2003; Padgett and Tabain, 2005; Smith, 2002; Steriade, 1995]. While there is still no agreement as to the precise nature of the relationship between these two phenomena, it is clear that the resolution of problems such as these is of crucial importance for the more general clarification of the relationship between phonetics and phonology, and in particular of questions concerning the role of phonetic information in formal theories of the phonological grammar. Among systems of phonological vowel reduction, Contemporary Standard Russian has held a position of particular prominence in the literature, in part because the system is relatively well described, but in part also because it is unusually complex as such systems go. (See Barnes [2002, 2006] and Crosswhite [2001, 2003] for typological surveys). What has made Russian interesting to phonologists in the past has been the apparent division of its reduction system into two distinct levels or degrees of reduction, one more extreme and one less so in effect. The present paper presents an experimental investigation of the facts of Russian vowel reduction, focusing in particular on the relationship between phonetic vowel durations and the 1 Throughout this paper, I will use the /.../ notation to indicate phonological representations, both “underlying” (Input) and “surface” (Output). I will use [...] to indicate phonetic transcription only. implementation of reduction patterns. Following up on work reported in Barnes [2006], wherein speech rate was manipulated to obtain a range of durational values for each target word, this study seeks to improve on those earlier results by eliciting both casual and hyperarticulated or clear speech versions of target words. I argue on the basis of the results obtained here that the Russian system of vowel reduction has in fact been mischaracterized traditionally: rather than having two patterns or degrees of phonological vowel reduction, the correct analysis of Russian includes only a single pattern of phonological vowel reduction, with what has traditionally been labelled Degree 2 reduction in fact being explicable entirely in terms of duration-dependent phonetic undershoot. This new analysis of the Russian facts is further argued to have important consequences not only for our understanding of the nature of vowel reduction, but also for our understanding of hyperarticulation as a process of enhancement of phonological contrast. In particular, consequences for the analysis of putative patterns of phonetic analogy or incomplete neutralization are explored. The remainder of this paper is structured as follows: section 2 provides background information on Russian vowel reduction and on the phonetic underpinnings of phonological vowel reduction more generally. Section 3 presents the results of the hyperarticulation experiment at the heart of the current study. Section 4 discusses these results in the broader context of vowel reduction, hyperarticulation and patterns of phonological neutralization. Section 5 sums up, drawing conclusions and pointing out several promising directions for the future. 2 2.1 Background Vowel reduction in Contemporary Standard Russian Contemporary Standard Russian is traditionally described as displaying two distinct patterns or degrees of vowel reduction [Avanesov, 1972; Bondarko, 1977; Halle, 1959; Jones and Ward, 1969; Lightner, 1972; Matusevic, 1976; Shcherba, 1912; Timberlake, 2004; Trubetzkoy, 1969; Ward, 1975; inter alia]. The first of these, henceforth Degree 1, is typically described as being less severe or extreme than the second, henceforth Degree 2. Degree 1 reduction is associated primarily (though see below for details) with the syllable immediately preceding the lexically stressed syllable, what I will call throughout the first pretonic syllable. The second degree of reduction is described as occurring in all other unstressed syllables. From the point of view of phonological contrast in the vowel system, the picture is as follows. Russian contrasts five vowels in stressed syllables, as shown in (1). (See Padgett [2001] and a long tradition of sources therein for arguments against a contrast between /i/ and the high central vowel //.) In unstressed syllables, the mid vowels [e] and [o] do not surface. In first pretonic syllables, the primary context for Degree 1 reduction, this results in the contraction of the vowel system to the three vowels [i, a, u] shown below, while in other unstressed syllables, Degree 2 reduction compresses this already-reduced inventory somewhat, its most apparent effect being the realization of [] in place of [a]. It should be noted, however, that even in Degree 1 reduction a certain amount of vertical compression of the vowel space is traditionally described as taking place, at least for some speakers: the vowel usually rendered as [a] in Western sources is in fact said to be realized somewhat higher, closer to IPA []. (The traditional transcription in Russian sources is [], which can be somewhat misleading, given this symbol’s IPA value.) Following (and particularly between) palatalized consonants, the effects of vowel reduction are more noticeable still, with the low vowel [a/] raising toward, and perhaps merging with [i]. (See Padgett and Tabain [2005], however, for some evidence against the completeness of this merger). (1) Stressed and Unstressed Vowel Inventories of Russian STRESSED SYLLABLES i DEGREE 1 REDUCTION u e i DEGREE 2 REDUCTION u o i u  a a In so-called “hard” or palatalization-free environments at least, then, expression of phonological contrast is affected in Russian vowel reduction solely through the exclusion of the mid vowels from unstressed syllables. In both Degrees 1 and 2 of reduction, for most contemporary speakers, underlying /e/ merges with /i/ (noting again the uncertainty here contributed by Padgett and Tabain), and underlying /o/ merges with /a/. It is important to note that the transition from Degree 1 reduction to Degree 2 does not actually reduce the number of contrasts available in the system. Instead, the difference between the two degrees of reduction lies most prominently in the degree of centralization of the reduced reflex of underlying /a/ and /o/. For this reason, it will be primarily these vowels that will be of concern in this paper. I will also focus exclusively on the realizations of these vowels in palatalization-free environments, acknowledging the difficult questions that this leaves unaddressed. The realization of /a, o/ in syllables in various positions relative to lexical stress is exemplified in (2), where in each case underlying /o, a/ surface as [o, a] under stress, merge as [a] in the first pretonic syllable, and raise to schwa in the second pretonic syllable. (2) STRESSED /o/ moldst ‘youth’ bol ‘pain’ /a/ starj razum 2.2 DEGREE 1 REDUCTION DEGREE 2 REDUCTION malodnkj ‘young’ (dim.) mladoj ‘young’ balet ‘to hurt’ blvoj ‘pain’ (adj.) ‘old’ (adj.) starik ‘reason’ razumn ‘old man’ strna ‘old times’ ‘wisely’ (adv.) rzumet ‘to understand’ Exceptional contexts At this point it is important to note that while vowel reduction in Russian is virtually exceptionless lexically (that is, no instance of lexical /e, o/ is permitted to surface as such in unstressed syllables), there are a number of contexts that are traditionally described as exceptions to the pattern of reduction described above. These exceptional contexts all involve the apparent failure of Degree 2 reduction to apply where it otherwise should (that is, anywhere but stressed and first pretonic syllables). Thus, what ought to surface as a schwa is instead realized closer to [a], or at least []. The first of these contexts is absolute word-initial position, as seen in (3). (3) a. b. /odnomu/ /obiajet/ [adnamu] [abiajt] ‘one’ (dat.sg.) ‘insult, abuse’ (3sg.) *[dnamu] *[biajt] Here lexical /o/ in absolute word-initial position fails to reduce to schwa, surfacing as [a/] instead. Note that this applies only to vowels in absolute word-initial position; /o, a/ in wordinitial syllables with onset consonants reduce as expected (cf. /podool/ ‘went’ (masc.sg.) ⇒ [pdaol]). The second exceptional context involves sequences of vowels in hiatus. According to Timberlake [2004: 51-2], when underlying /a/ or /o/ is the second of (any) two vowels in sequence, or when it appears immediately before another instance of underlying /a/ or /o/, raising to schwa fails to apply (or applies less dramatically). (4) shows examples of /o, a/ in second and third pretonic syllables failing to reduce to schwa in the relevant contexts. (4) a. b. /odnoobraznj/ /sootnoenije/ [adnaabraznj] [saatnaenije] ‘monotonous’ ‘relationship’ *[adnabraznj] *[satnaenije] The last exceptional context, less commonly cited in the literature than the two previous, is absolute phrase-final position. When appearing in phrase-final position, lexical /o, a/ fail to reduce to schwa, or do so only “partially” or “optionally” [Matusevic, 1976: 102; Zlatoustova, 1962: 109-139]. Full Degree 2 reduction is avoided. These exceptional contexts for Russian vowel reduction will become important in the arguments presented below. 2.3 Past Analyses Traditional analyses of Russian vowel reduction do not draw a distinction of the sort assumed today by many phonologists between abstract, symbolic rules or processes affecting phonological representations and the phonetic implementation of those representations. In most cases, therefore, the question as to whether the two “degrees” of vowel reduction both apply at the same level of representation is simply never addressed explicitly. Degree 1 and Degree 2 reduction are nonetheless clearly portrayed as the same kinds of processes, applying simply to a greater or a lesser extent. Trubetzkoy, for example, notes the complementary distribution of Degree 2 [] and Degree 1 [a/], and describes them as positional variants of the phonemes /a/ and /o/, which are neutralized in favor of the low unround vowel in unstressed syllables. In the generative phonological tradition (arguably beginning in both the local and general senses with Halle [1959], and further exemplified by Lightner [1972] and Alderete [1995], among others), the traditional analysis of Degree 1 and Degree 2 reduction is continued largely intact, to the extent that in both cases phonological rewrite rules (or more recently the interaction of constraints on phonological representations) replace underlying phonological /a/ or /o/ with appropriate output forms /a/ or // as the case may warrant. Importantly, both patterns are considered to be equally “phonological” in nature. Any distinction between the two is only a matter of degree, rather than of kind. Crosswhite [2001] represents an important break with this tradition. For Crosswhite too, both Degree 1 and Degree 2 reduction are phonological processes. Where they differ is in their functional motivations (and hence for Crosswhite also in their formal implementations). Degree 1 reduction is an example of what Crosswhite calls “contrast-enhancing” vowel reduction. The idea is that in contexts where discriminating vowel contrasts might be difficult (such as unstressed syllables), constraints against the realization of certain marked vowels have the effect of creating greater perceptual distance between the vowels that are ultimately allowed to surface. Thus, mid vowels are banished from unstressed syllables, while the high and low vowels are permitted. (This account assumes that first pretonic [a/] is in fact a low vowel.) Crosswhite singles out these Degree 1 environments as distinct from Degree 2 environments structurally by assuming, with Halle and Vergnaud [1987], that the first pretonic syllable and the tonic together constitute the head foot of the word. (Stress in Russian is lexically contrastive, culminative, and obligatory; secondary stresses are unattested in non-compound forms.) All other syllables are analyzed as unfooted, weaker both prosodically and in terms of licensing capacity for vowel contrasts. Degree 2 reduction is thus analyzed by Crosswhite as something altogether different in motivation from Degree 1 reduction, something that she calls “prominence-reducing” reduction. In this analysis, vowels outside the head foot in Russian as taken to be non-moraic, moras being assigned only to vowels located in footed syllables. There is, furthermore, a constraint against the realization of relatively more sonorous vowel qualities (i.e. /a/) in non-moraic contexts.2 The formal implementation of this ban is less important in the present context than the simple fact that the constraints involved ultimately necessitate the substitution of the less sonorous vowel // for the /a/ that might otherwise have surfaced in these positions. While the idea of reducing the sonority of an unstressed vowel does seem related to the notion of gestural undershoot, Crosswhite is careful to distinguish phonological prominence-reduction from phonetic undershoot. While the latter is a change in vowel quality resulting from a decrease in duration, the former is a decrease in duration resulting from a change in vowel quality [Crosswhite 2001: 46]. 2.4 Vowel reduction in phonetics and phonology Crosswhite is not alone, of course, in noting the similarity between some phonological vowel reduction patterns and the kind of phonetic undershoot originally attributed to decreased vowel durations by Lindblom [1963]. Flemming [1995/2002, 2001, 2005] and Barnes [2002, 2006] both posit a causal link of some kind between duration-dependent undershoot and patterns of phonological vowel reduction. Such a link makes intuitive sense for a number of reasons: It is well-known that lower vowels have longer durations, all things being equal, than higher vowels [Lehiste, 1970]. This difference is typically attributed to the additional time it takes to reach the more open targets for the articulators of the upper vocal tract characteristic of lower vowels. It is also the case that most languages with phonological vowel reduction exhibit large differences in duration between lexically stressed and unstressed vowels, unstressed vowels undergoing severe shortening in comparison to their stressed counterparts. (Exceptions to this tendency do exist. Shimakonde [Liphola 2001], for example, removes mid vowels from unstressed syllables even when they are bimoraic. See section 4.4 below for more on patterns such as this.) Within the Slavic family, Russian (see below), Belarusian [Czekman and Smułkowa, 1988: 226], and Bulgarian [Tilkov, 1982] all have lexically contrastive stress cued in part by severe shortening of most unstressed vowels, and all three also have phonological vowel reduction; on the other hand, Polish [Jassem, 1959], Czech [Lehiste, 1970: 36-7], and Macedonian [Savitska and Spasov, 1997: 172] all have fixed stress not associated with such a degree of shortening, and they also 2 Note that in unifying Degree 2 environments as “unfooted syllables”, this analysis encounters difficulties accounting for the failure of Degree 2 reduction in the exceptional contexts discussed above, since none of those exceptional syllables are obviously “footed”. Absolute initial vowels, for example, are treated with a constraint stipulating that absolute initial vowels bear a mora. Another constraint is posited banning schwa in hiatus. Absolute phrase-final vowels are not treated, but presumably could also be stipulated to bear moras along with absolute wordinitial vowels. lack phonological vowel reduction (though not phonetic vowel reduction, on which see Nowak [2006] for a detailed study of the problem in Polish). Likewise, in Serbo-Croatian, though stress is not fixed, and there are significant differences in the durations of stressed and unstressed vowels, nonetheless unstressed vowels fail to shorten to the extremes we find in languages like Russian and Bulgarian. Lehiste and Ivić [1986: 69-70] give average durations for unstressed short vowels in Ivić’s speech in a variety of prosodic contexts typically falling somewhere between 70 and 100 ms., well above the mean durations we will see below for most unstressed syllables of Russian; and as predicted, Serbo-Croatian does not exhibit phonological vowel reduction. Crosswhite [2001] and Barnes [2002] both demonstrate from broad cross-linguistic surveys that phonological vowel reduction overwhelmingly targets vowel height contrasts (rather than, say, contrasts of palatality, rounding, or ATR). Furthermore, when height contrasts are neutralized, it is invariably the low and/or mid vowels whose realizations are affected; high vowels remain effectively unaltered. Thus, the stressed and unstressed inventories shown in (5), with elimination of underlying mid and low vowels in favor of high vowels and schwa in unstressed syllables, is cross-linguistically common (e.g., Russian second pretonic syllables in nonpalatalized contexts). The stressed-to-unstressed mapping in (6), on the other hand, in which high vowels are prohibited in unstressed syllables while mid and low vowels persist, is unattested. Likewise unattested is the mapping in (7), in which three levels of vowel height remain contrastive, while the front/back contrast is eliminated. There are, in other words, no vertical vowel systems that are restricted to unstressed syllables, and no vowel reduction analogues to the neutralizations of palatality contrasts found in the vowel harmony systems of many Turkic or Uralic languages.3 (5) STRESSED UNSTRESSED i u e o i u  a (6) STRESSED UNSTRESSED i u e o e a (7) 3 o a STRESSED UNSTRESSED i u  e o  Though see Barnes [2006] on the reduction system of Sosva Mansi, and the possibility of reduction systems developing over time out of earlier harmony systems. a a Barnes and Flemming both seek to explain these patterns as the result of duration-dependent undershoot of F1 targets for mid and low vowels, resulting in a vertical compression of the vowel space. In this compressed vowel space, height contrasts would become less robust perceptually, with contrasting vowels now closer to one another than typically are their stressed analogues. Flemming [2005] presents a model deriving optimal unstressed vowel inventories using the principles of Adaptive Dispersion originally developed in Liljencrants and Lindblom [1972]. Flemming takes as his model for the effects of undershoot the output of the equations used by Lindblom [1963] to model the results of his one-speaker study of eight Swedish vowels. In this study, the relevant vowels were situated in nonsense words of the form /bVb/, /dVd/ and /gVg/. Each resulting monosyllable was embedded in four sentence frames differing in the target word’s sentence position and (phrasal) accentedness, such that a relatively wide variety of vowel durations (80-300 ms) were ultimately obtained. In his analysis, Lindblom found the major predictors of formant values for any given vowel to be duration and distance from consonantal locus to vowel formant target. He takes this result to argue for a view of undershoot as coarticulation, rather than the uniform centralization that others [e.g., Tiffany, 1959] had previously assumed. For F2, of course, this means that the direction and degree of undershoot would change as a function of both consonantal context and the F2 target value for the vowel in question. For F1, however, this result means something different. Lindblom’s model of F1 undershoot predicts no undershoot at all for vowels with target F1 below 375 Hz (meaning, in effect, that high vowels are unaffected). For vowels with F1 targets above 375 Hz, Lindblom models F1 undershoot as an exponential function of decreasing vowel duration with an equation similar to those used to model F2 undershoot. The general form of the F2 equations was αe−βt + F1(target), where t represents vowel duration, α is a constant proportional to the distance between consonant locus F2 and target F2 for each vowel, and β is a constant fixed for each consonantal context. In the F2 model, then, α could vary in both sign and magnitude as a function of F2 consonant locus relative to vowel target, meaning that F2 could either raise or lower to varying degrees as a function of undershoot. To the extent that all consonant places can be seen to have F1 loci lower than target values for any vowel, in the F1 model, all consonants affected the realization of vowels in the same fashion, proportionally to the difference between the vowel's F1 target and 375. Since only vowels with targets greater than 375 were submitted to this equation in Lindblom’s model, α had the same sign for all vowels and all consonants. In other words, while the degree of undershoot a given vowel would experience might change from context to context, the direction of undershoot was always the same: mid and low vowels raise. The parallelism between this phonetic effect, identified in the realization of Swedish stressed vowels of varying durations, and the cross-linguistic typology of phonological unstressed vowel reduction, is striking indeed, and strongly suggestive of a relationship of some sort between the two phenomena. Indeed, Flemming [2005] characterizes the undershoot obtained in Lindblom’s study as causing primarily vertical compression of the vowel space, and argues that this is the ultimate source of the height bias in the cross-linguistic typology of phonological vowel reduction.4 For Flemming, 4 It is interesting, incidentally, that this is not how Lindblom himself summarizes his findings. In fact, Lindblom characterizes the undershoot he obtains in his study as affecting primarily F2, while “in most cases there is little the causal connection between undershoot and phonological vowel reduction exists synchronically in the grammars of phonological vowel reduction languages. The essence of this proposal involves the interplay of duration-dependent undershoot and the principle of Adaptive Dispersion [Liljencrants and Lindblom, 1972], operationalized by Flemming in an Optimalitytheoretic framework [Prince and Smolensky, 1993/2004] assuming a unified model of phonetics and phonology. The idea is that at sufficiently decreased vowel durations, there is no way for speakers to maintain adequate perceptual distance between vowels in the inventory without expending unacceptable amounts of articulatory effort (an additional factor conditioning undershoot added to the earlier Lindblom model by Moon [1991], and Moon and Lindblom [1994]). In this model, factors such as vowel duration and articulatory effort interact with constraints mandating the maintenance of phonological contrasts, with the result in this case that speakers prefer to abandon certain contrasts altogether, rather than expend the biomechanical energy necessary to maintain them. The result is fewer available vowel contrasts in unstressed positions within an optimally dispersed, albeit smaller, vowel inventory. This type of approach to phonological vowel reduction, relying as it does on shortened vowel durations and substantial, active synchronic formant undershoot for the derivation of phonological neutralization, succeeds in expressing formally the link between undershoot and neutralization, and in making robust and quantitatively explicit typological predictions about the sizes and shapes of unstressed vowel inventories. It does so, however, at the cost of making undershoot a necessary and (under certain assumptions) sufficient condition for phonological vowel reduction. The problems that arise from this move are discussed in greater detail in section 4.4 below. By contrast, Barnes [2002, 2006] argues that the link between undershoot and phonological vowel reduction is instead a diachronic one. As in the synchronic model, compression of the vowel space along the F1 dimension due to undershoot makes vowel height contrasts less robust perceptually. Adopting the listener-oriented approach to sound change pioneered by John Ohala [passim], the diachronic approach holds that the increased confusability caused by substantial undershoot makes it more likely that the relevant contrasts will in fact be misperceived frequently in the course of actual language use. Consistent misperception of height contrasts in unstressed syllables leads to their ultimate collapse, as listeners/learners reinterpret the effects of undershoot as speakers’ intended target realizations for the vowels in question. A sound change thus takes place whereby previously contrasting vowels that had drifted too close to one another in perceptual space merge into a single phonological category in unstressed syllables. The Bulgarian dialects mentioned in section 1 illustrate this nicely, with modern-day dialect geography recapitulating in space the historical development of the dialects through time. undershoot in F10 observable even at short durations of the vowel...” [1963: 1776]. Given this, we might wonder why F2 undershoot seemingly never results in neutralization of phonological contrasts in the way that F1 undershoot so often does. Padgett and Tabain [2005] discuss this problem in some detail, positing as one explanation the inconsistency in both direction and degree of F2 undershoot in different consonantal contexts, compared to the monolithic raising effect of F1 undershoot across contexts. Another factor that may be important to bear in mind is that Lindblom's model is based on results that included no vowels under 80 ms in duration. Unstressed vowels in vowel reduction languages are substantially shorter than this (cf. the mean durations for Russian presented in Table 3 and Table 4 of the present study, and the results of the speech rate study of Barnes [2006]). This leaves open the possibility that at the relevant durations undershoot of F1 increases more dramatically than does that of F2, in a way that would necessitate modification of Lindblom’s original equations. The Western dialects of Bulgarian (including the native dialect of the region surrounding the capital, Sofia) represent the first stage in this historical progression, reportedly lacking reduction of unstressed vowels of any kind.5 Thus /darove/, ‘gifts’, /zeleno/, ‘green’ (neut.sg.), and /bata/, ‘father’, would be pronounced [darove], [zeleno], and [bata] respectively [Stojkov, 1993: 144]. In (most of) these dialects, then, there are six vowels /i, e, a, , o, u/ contrasted in both stressed and unstressed syllables. At the opposite end of the spectrum, many Eastern dialects, for example the Central Balkan dialects spoken around Lovec, Trojan, and Gabrovo, have a complete merger of /i-e/, /-a/, and /u-o/ in unstressed syllables [Stojkov, 1993: 109]. Thus, in contrast to the Western examples given above, here we find /darove/, ‘gifts’ pronounced [druve], /zeleno/, ‘green’ (neut.sg.), pronounced [zilenu], and /bata/, ‘father’, pronounced. [bta]. The stressed and unstressed vowel inventories of these dialects are given in (8). (8) Phonological vowel reduction in Central Balkan Dialects of Bulgarian STRESSED i e UNSTRESSED u  o i ⇒ u  a The putative link between these two systems, one with phonological vowel reduction, and one without, are dialects in which the mid and low vowels are raised phonetically toward, but do not yet merge with, their higher counterparts in the system. Stojkov [1993: 211] identifies the Rupic dialects as displaying such a system, which he calls “half-reduction” (polureduktsija). (Rupic is a cover term for most of the dialects of southeastern Bulgaria, including Thrace and the Rhodopes. By no means, however, do all these dialects share the same system of vowel reduction.) In the relevant dialects, /darove/, 'gifts', /zeleno/, 'green' (neut.sg.), and /bata/, 'father', would be pronounced something like [darove], [zeleno], and [bata]. The diachronic progression through these three stages of development, going from an initial state with no appreciable reduction, phonetic or phonological, to substantial phonetic reduction resulting in vertical crowding of the vowel space, and finally to collapse of those perceptually compromised contrasts, is illustrated in schematically (9) below.6 (9) Development of unstressed vowel inventories in Bulgarian dialects 5 This claim is surely too strong in its most literal sense; some formant undershoot no doubt occurs if there are even minor durational differences between stressed and unstressed vowels (which there almost certainly are, given the documented presence of such differences in Serbo-Croatian [Lehiste and Ivić, 1986] and Macedonian [Savitska and Spasov, 1997], the languages to which the various Western dialects of Bulgarian are ultimately transitional). Traditional descriptions of the mid and low vowels of these dialects, in any case, count unstressed /e, a, o/ as “not differing” from their stressed analogues [Stojkov 1993: 144], where we might take this just to mean that any existing difference is perceptually insignificant. 6 A complete picture of this progression will be somewhat more complex. It is likely in particular, given the facts in other dialects, that the various mergers between stages II and III themselves occurred in stages, with the merger of /a/ and schwa happening first, followed by that of /o/ and /u/, and only lastly by /e/ and /i/. See also on this Scatton [1983: 55-57]. Sadly, accounts of the phonetics of vowel reduction in Bulgarian dialects are still almost exclusively impressionistic in nature. STAGE I: NON-REDUCTION i e u  o STAGE II: PHONETIC COMPRESSION i e  a u o STAGE III. MERGER i u  a It is not the purpose of the present study to adjudicate definitively between the diachronic and synchronic approaches to the relationship between phonetic undershoot and phonological vowel reduction. Rather, for present purposes, it is enough to recognize that in both approaches, duration-dependent undershoot is assigned a causal role in the formation of patterns of phonological vowel reduction. 2.5 Undershoot and Russian vowel reduction Given the hypothesized link between durational impoverishment of unstressed syllables and systems of phonological vowel reduction, Barnes [2006] designed an experiment to determine the relationship between vowel quality and vowel duration in the reduction system of Modern Russian. This experiment focused in particular on first formant values for underlying /a/ and /o/ in stressed, first pretonic, and second pretonic syllables. In order to achieve a range of differing vowel durations for a given set of strings, subjects were asked to read through a list of target utterances twice, once at a comfortable, conversational tempo, and then a second time as fast as possible. Target strings were 72 real trisyllabic words of Russian, in which the relevant vowels were all located in open syllables followed by unpalatalized voiced or voiceless stops and preceded by unpalatalized voiced stops, voiceless stops, or laterals. These segment types were chosen primarily to facilitate segmentation for duration measurements. Target forms were embedded in frame sentences of the form Mashka X skazala, ‘Mashka said X’, and each target word appeared twice for a total of 144 sentences. Sentences were randomized and arranged in blocks of 12. Four subjects participated in the study, details concerning whom are presented below in section 3.1.1. Target word types from this study are exemplified in (10). (10) Sample target words from the speech rate study reported in Barnes [2006] a. Stressed /a/: b. Stressed /o/: c. First pretonic /a/: /lopata/ /pooda/ /napadat/ d. First pretonic /o/: /olodat/ e. Second pretonic /a/: /takovoj/ f. Second pretonic /o/: /odovoj/ [lpat] [pod] [npdat] (cf. [napatki] [ldat] (cf. [lodnj] [tkvoj] (cf. [tak] [dvoj] (cf. [got] ‘shovel’ ‘weather’ ‘attack’ (inf.) ‘attacks’ [nom.pl.]) ‘starve’ (inf.) ‘hungry’ [masc.sg.]) ‘such’ (masc.nom.sg.) ‘thus, so’) ‘annual’ (masc.nom.sg.) ‘year’) Major findings of this study were as follows: First, in line with all previous descriptions, the neutralization of /a/ and /o/ in unstressed syllables was found to be categorical (at least in terms of first formant and durational values, the only variables controlled in this experiment). No significant difference between underlying phonological categories was preserved in unstressed position. Mean values and standard deviations for these measures for all speakers and positions are given in Table 1 and Table 2. More surprising was the relationship between F1 values and durational measures for the three positions. The first point concerns the means. The relatively low mean durations for second pretonic /a/ and /o/ seem to line up well with the schwa-like F1 values of the relevant vowels; to the extent that reduction, and reduction to schwa in particular, is in any sense a product of duration-dependent undershoot, these values are to be expected. Problematic in this light, however, are the durational values for first pretonic and stressed /a/ and /o/: Mean duration for first pretonic /a/ and /o/ was somewhat shorter than that of stressed /a/ (69 and 70 ms vs. 82 ms respectively). However, even if we believe (and we do not) that a mean durational reduction of 12 or 13 milliseconds (15 or 16%) could in principle cause enough undershoot that an otherwise robustly distinguishable [a] and [o] might be forced unacceptably close to one another, mean F1 values show that in this study at least, this has not in fact occurred. If anything, mean F1 for first pretonic /a/ is actually slightly higher than it was for stressed /a/. In other words, whatever durational impoverishment there was here has not caused the low vowel in the system to raise at all. This result represents a serious challenge to any approach to vowel reduction that seeks to derive neutralization synchronically from facts about phonetic duration. The contrast between Russian /a/ and /o/ is neutralized in the first pretonic syllable despite the fact that little durational impoverishment, and no vowel space compression, have taken place. In Barnes [2006], I argue that this is because neutralization of /a/ and /o/ in Russian unstressed syllables is a categorical phonological process triggered by abstract structural factors, in particular presence vs. absence of lexical stress. Neutralization is thus phonetically unmotivated, not to say unnatural, in the Russian first pretonic syllable. See Barnes [2006] for potential diachronic accounts of this pattern, and see section 4.4 below for further discussion of the implications of this finding. [Table 1 about here] [Table 2 about here] Looking more closely at the data summarized above yields a second important observation. The scatterplot in Figure 1 shows F1 measurements for second pretonic /a/ and /o/ for all speakers as a function of vowel duration, and it is here that the relevance of the manipulation of speech rate begins to be apparent. What we expect in second pretonic position from phonological descriptions of Russian is Degree 2 reduction (i.e. []). What we in fact find is substantially more nuanced. While the mean given above suggests a schwa-like vowel, closer inspection shows a strong positive correlation (r = .532, p < .0021, N = 357) between F1 values and vowel duration: the longer the vowel got, the lower it was realized, and conversely, the shorter it was, the higher. In other words, Degree 2 reduction appears from this data to apply gradiently, with decreases in duration yielding more reduction and increases yielding less. As much then as the neutralization of /a/ and /o/ is oblivious to phonetics, the raising of /a/ to schwa appears not just to be phonetically-motivated in the loose sense in which this term is usually employed in phonology, but actually phonetically-driven, in the sense of direct causation in performance. [Figure 1 about here] The difference between categorical and gradient application of processes has been taken by many to be a defining distinction between symbolic phonological processes and quantitative phonetic processes [Barnes, 2002; Cohn, 1990; Keating, 1988, 1996; Myers, 2000; Zsiga, 1993; inter alia]. Assuming the validity of this distinction, Barnes [2006] concludes that so-called Degree 2 reduction in Russian is best treated as a matter of phonetic implementation. This means that, contrary to all previous reports, Russian has not two, but one phonological vowel reduction pattern: the neutralization of /a/ and /o/ in all unstressed syllables, with substitution of /a/ for underlying /o/. In positions such as stressed syllables or first pretonic syllables, positions where vowel durations tend to be longer, phonological /a/ is realized phonetically as a low vowel, [a/]. Where durations are shorter, as in most other unstressed syllables, low target articulations for /a/ are subject to undershoot, with the result that the low vowel is raised phonetically in proportion to the reduction in its duration in any given instance. There is no evidence that the phonology of Russian has any role in (or access to) this process of gradient duration-dependent raising. In terms of phonological categories, then, there is in fact no schwa in Russian; there is only /a/, which may be undershot to varying degrees as a function of duration.7 If this analysis is correct, then our account of Russian phonology is simplified dramatically, in a way that requires only the independently attested phenomenon of gestural undershoot as deployed in a language-specific phonetics. The payoff increases, however, when we consider the “exceptional contexts” discussed above in section 2.2, where raising to schwa fails to apply. This set of contexts, again, is absolute word-initial position, various configurations in hiatus, and absolute phrase-final position. In terms of phonological structure, this set of positions is disparate indeed, and for the phonological account of Degree 2 reduction to succeed, it is forced to posit three architecturally unrelated reasons for the failure of Degree 2 reduction in each context. If Degree 2 reduction is in fact just duration-dependent undershoot, however, a single, unified account of the exceptional contexts emerges with no elaboration of the analysis: all three exceptional positions are associated with increased phonetic segment durations; indeed, crosslinguistically all three are associated with resistance to reduction processes of precisely the type we find in Russian. Absolute word-initial position is associated cross-linguistically with articulatory strengthening and/or durational enhancement of consonants and vowels, the details of which are the subject of an extensive experimental literature [Byrd, 2000; Byrd, Krivokapić, and Lee, 2006; Byrd and Saltzman, 2003; Cho, 2002; Cho, 2005; Cho and Jun, 2000; Cho and Keating, 1999; Cho and McQueen, 2005; Cho, McQueen, and Cox, in press; Dilley, Shattuck-Hufnagel, and Ostendorf, 1996; Fougeron and Keating, 1997; Keating, Cho, Fougeron and Hsu 1999, Oller 1973]. Other languages with sound patterns involving absolute word-initial vowels, the effects of which could be derived synchronically or diachronically from initial strengthening include Nawuri [Casali, 1995; Kirchner, 1998], Luganda [Hubbard, 1994: 161-165] and Runyambo [Larry Hyman, p.c.]. The additional segment duration and increased gestural magnitudes that have been found in absolute word-initial position in a variety of languages could easily serve to prevent the application of Russian Degree 2 reduction, since either, or more likely both together, would have the effect of preventing undershoot in the production of low vowels. Absolute phrase-final position is also associated cross-linguistically with additional duration for vowels, a phenomenon known as domain-final lengthening [Beckman and Edwards, 1987; 7 In Optimality-theoretic terms, what this means is that the Markedness constraint against whatever constellation of features we take to represent schwa (call it *Schwa for short) outranks all relevant Faithfulness constraints. Phonological schwa is not tolerated as a representation in any position, stressed or unstressed. Beckman, Edwards, and Fletcher, 1991; Byrd, Krivokapić, and Lee, 2006; Byrd and Salzman, 2003; Cambier-Langeveld, 1997, 1999; Keating, Wright, and Zhang 1999; Klatt, 1975; Oller, 1973; Wightman, et al., 1992; inter alia]. The relationship between this increased duration and gestural magnitude is still somewhat unclear, and certain cases have been identified in which final lengthening demonstrably fails to undo the effects of undershoot in final position. (See Johnson and Martin [2001] for experimental details on Creek, and Barnes [2006: 141-151] for a crosslinguistic survey.). However, there is also abundant evidence that final syllable vowels, and especially phrase-final vowels, are resistant to a variety of reduction or assimilation processes which should otherwise target them in a wide array of languages, including Belarusian, Ukrainian dialects, Brazilian Portuguese, Eastern Mari, Uyghur, Hausa, Catalan dialects, English, Yakan, Maltese, Nawuri, Shimakonde, Bonggi, and Timugon Murut [Barnes 2006: 79114].8 Russian fits this pattern nicely, and to the extent that Degree 2 reduction is the result of compressed segment durations, it would be unsurprising to see it circumvented by the increased durations associated with phrase-final position. The last exceptional context, hiatus before [a], is similarly explicable. If, again, reduction of [a] to schwa is in fact a consequence of duration-dependent undershoot, a sequence of two like vowels uninterrupted by other gestures should certainly provide adequate duration for articulators to reach their targets, effectively undoing Degree 2 reduction. The problem of exceptional contexts is discussed further in section 4.3, together with suggestions for next steps in understanding resistance to reduction in these three positions. There was, however, a problem. The two crucial experimental findings supporting the analysis described above were the apparent irrelevance of speech rate or duration to the application of neutralization of /a/ and /o/ in unstressed syllables, and the strong correlation between extent of raising toward schwa and vowel duration. The prediction that this analysis makes is that, given sufficient segmental durations (at, for example, an extremely slow rate of speech), reduction toward schwa should fail across the board. In other words, we should find no difference between vowel qualities in second pretonic and first pretonic syllables. The speech rate study of Barnes [2006] did not, however, succeed in demonstrating this conclusively. As can be seen in Figure 2, while segment durations (and hence first formant values) did approach values recorded for the shortest first pretonic tokens of /a/ and /o/, there was almost no overlap in the realizations of the vowels in these two positions. Barnes [2006] takes this lack of overlap to reflect constraints on the durations of second and first pretonic vowels in normal speech: apparently, even in the slowest normal speech, second pretonic vowels will only lengthen a certain amount, with an upper bound somewhere around 50 to 60 milliseconds, and this turned out not to be enough to completely avert undershoot. The prediction, however, is that, had a fuller range of durational values been obtained, we would have seen not two distinct clouds of productions as in the scatterplots shown here, but rather a single curve that grew at first steeply, and then more gradually as it approached an asymptote representing target F1 for the Russian vowel [a]. [Figure 2 about here] One other aspect of the scatterplots in Figure 2 is important to note before moving on. While it is true that second pretonic /a/ and /o/ never lengthened into the range of durational values characteristic of their first pretonic or stressed syllable analogues, it is not the case that we see no durational overlap at all between the two categories. For one speaker in particular in fact, DT, first pretonic vowels produced at the faster tempo did get short enough in a number of instances that we actually see substantial overlap between the shortest first pretonic /a/ and /o/ and the longest second pretonic vowels. Interestingly, however, F1 values for DT's first pretonic vowels stayed consistently above those of the second pretonic, as is immediately visible from the scatterplot. In other words, the data from this speaker would be more accurately modeled not by one curve, but by two. This is odd from the point of view of the Russian data in the present study, but nonetheless interestingly reminiscent of observations made about sources of phonetic target undershoot in languages such as English, Swedish, and Dutch [e.g., Engstrand, 1988; Kuehn and Moll, 1976; Moon and Lindblom, 1994; Nord, 1975, 1987; Van Son and Pols, 1990, 1992]. Though the details and conclusions of these studies differ greatly, they all present evidence of one sort or another suggesting duration alone is insufficient to account for patterns of target undershoot. Particularly relevant is the work of Moon and Lindblom [1994] on phonetic vowel reduction and clear speech in English. In this study of undershoot of F2 targets, Moon and Lindblom independently varied vowel duration (induced through polysyllabic shortening) and degree of hyperarticulation, a binary contrast between what they call Citation Form, or ordinary pronunciation of forms in isolation, and a level of Clear Speech obtained by instructing subjects to “overarticulate”. While duration-dependent undershoot was observed (at least for lax vowels) in both registers, in many cases duration alone was insufficient to predict degree of undershoot; specifically, for a given duration, vowels in the Clear Speech register were subject to less undershoot than vowels in the Citation Form register. In this English data, then, hyperarticulation of vowel quality seems to involve changes in some other variable in addition to increased vowel duration. This other variable, in the context of the H&H model of Lindblom [1990], is hypothesized to be articulatory effort, signalled by an increase in movement velocities, allowing a more distant target articulation to be achieved in a fixed amount of time. These increased velocities are achieved through an increase in force amplitudes for the relevant gestures, which in turn would require an increase in “biomechanical power expenditure”. Target undershoot can be avoided, in other words, either by increasing the duration allotted for the achievement of target values, or alternatively, by increasing the amount of effort expended in order to achieve those same values in a given allotment of time. Given these results, it is tempting to see DT's productions in the speech rate study of Barnes [2006] in a similar light. For DT, while increased duration in second pretonic syllables did result in less raising of phonological low vowels toward schwa, corresponding decreases in the durations of first pretonic vowels did not guarantee analogous increases in the degree of undershoot the unstressed vowel was subject to. Rather, first pretonic /a/ retained its higher F1 values even in the face of substantial reductions in duration, suggesting that for DT preservation of target F1 values in first pretonic syllables is important enough to warrant the expenditure of sufficient articulatory effort to stave off the effects of undershoot. This strategy of DT's becomes relevant again in the discussion of outliers in the hyperarticulation experiment presented in the following section. The failure of the manipulation of speech rate to produce greater overlap in durational values for second pretonic and first pretonic syllables in this study makes it difficult to determine whether other speakers share DT's strategy for the preservation of first pretonic F1 target values. It also makes it impossible to know the full extent of the curve relating duration to F1 values in second pretonic syllables: while duration and F1 values here approach those of first pretonic syllables, it is not certain that given sufficient duration, they would come to overlap with them as well. In order to remedy this problem, a further experiment was subsequently devised, which constitutes the heart of this report.