The Experiment - Hyperarticulation And Vowel Reduction In Russian

3 The Experiment: Hyperarticulation and vowel reduction in Russian While Barnes [2006] did show that F1 values for Degree 2 reduced /a/ and /o/ increased as a function of the increased duration they received at slower speech rates, even the slowest speech rates were insufficient to produce durational values for these vowels that would result in canonical [a]-like quality. It seems from this that Russian places an upper limit on the duration of Degree 2 vowels in normal speech that keeps them regularly somewhat shorter than phonemically analogous vowels in the first pretonic syllable. The result of this was clear in the scatterplots displayed above: effectively missing was crucial data from the area between the canonical durational values for Degree 1 and Degree 2 of reduction; a gradual transition between the two levels of reduction is suggested, but not definitively displayed, by the data collected there. The present experiment was devised to remedy this shortcoming. If the durations of Degree 2 vowels are severely limited under normal circumstances, these limitations seem to fall away under conditions of extreme clear speech or hyperarticulation. This effect was known to me impressionistically through informal observation of Russian speakers’ productions in contexts demanding hyperarticulation. A sizable literature on clear speech and hyperarticulation, albeit dealing almost exclusively with data from English, suggested that this would be the case as well. This literature is focused primarily either on the extent to which clear speech enhances intelligibility for a given population or communicative context, or on the features that are enhanced to achieve any putative added intelligibility. A number of works have documented either increased segment durations, or an expanded vowel space (or both) as characteristics of clear speech in some language [e.g., Bradlow, et al., 2003; Ferguson 2003; Liu, et al., 2004; Picheny, et al., 1986; Uchanski, et al., 1996]. Of particular interest is the work of Moon and Lindblom [1994] introduced above, demonstrating in English a mitigation of duration-dependent undershoot in clear speech too substantial to be attributed to duration alone, suggesting the activity of an additional variable, articulatory effort, in determining vowel realization. Lastly, in a much-needed extension of the study of clear speech to a language other than English, Smiljanić and Bradlow [2005] find essentially the same durational enhancement and vowel space expansion in both English and Croatian, despite substantial differences in vowel inventory size between the two languages. All this suggests that the elicitation of clear speech or hyperarticulation would be an effective method of overcoming the problems encountered in Barnes [2006]. A situation was thus constructed in which hyperarticulated versions of a set of target words would be produced, with the aim of attaining vowels with durational values long enough in principle to license any vowel quality a speaker might wish to produce. As reported in the following sections, this approach was largely successful. 3.1 3.1.1 Methodology Speakers The same four native speakers of Contemporary Standard Russian participated in this study as were included in Barnes [2006]. These were three females (RR, from Ufa, TM, from Saint Petersburg, M, from Moscow) and one male (DT, from Moscow). All had been born, raised, and educated through the secondary level in a predominantly Russian-speaking environment, and all were between the ages of 18 and 35 at the time of recording. All reported continued frequent use of Russian here in the United States with family and friends. Three of the four speakers had some level of training in linguistics, though all were naive as to the purpose of the experiment. 3.1.2 Stimuli Stimuli were 18 real trisyllabic words of Russian. All words had lexical stress on the final syllable, and the vowel [] (unstressed /a/ or /o/) in the first pretonic syllable. In the target, second pretonic syllable, 9 of the words had underlying /a/, while the other 9 had underlying /o/. All target syllables were open, and all target vowels were followed by unpalatalized voiced or voiceless stops and preceded by unpalatalized voiced stops, voiceless stops, or laterals. These limitations were chosen primarily to facilitate segmentation and formant analysis later on. Examples are shown in (11) below, where the relevant vowels are bolded in IPA transcription. The complete list is given in the Appendix. It should be noted that words were included in the “underlying /o/” category only when the relevant morpheme surfaces under stress, and hence as phonetic [o], in a morphologically related form. Many words of Russian contain vowels [] or [] that are spelled with orthographic ‘o’, but which never surface under stress in the language, such that evidence for their underlying status as /o/ is questionable. While no such words were included in the /o/ category here, several of the /a/ words tested did in fact lack morphologically related stressed [a] counterparts (e.g., kabare, [kbre], ‘cabaret’).9 It was felt, however, that in this instance orthographic evidence (a-spellings), together with [a]-like surface realizations, were sufficient to support their inclusion in the /a/ category (that is, rather than their inclusion in the /o/-category, the only other possibility, and one for which there is no evidence of any kind). (11) UNDERLYING /a/ /takovoj/ [tkvoj] cf. /tak/, [tak], ‘thus, so’ ‘such’ (adj.) UNDERLYING /o/ /bokovoj/ [bkvoj] cf. /bok/, [bok], ‘side’ ‘side’ (adj.) These target words were placed in frame sentences of the form Po-moemu on X skazal, ‘It seems to me that he said X’. These sentences were presented to subjects in the form of a printed list, written in Cyrillic orthography, in pseudo-random order. To achieve hyperarticulated versions of the target words, an experimental paradigm used in, e.g., Johnson, Flemming and Wright [1993], was adapted to these circumstances. Speakers were asked to read through the list of sentences at a comfortable, conversational rate of speech. After each sentence, however, speakers were asked twice to repeat themselves by the experimenter, simulating incomprehension. (Speakers were informed before the experiment began that this would occur, though apparently advance warning did little to make the experience less irritating.) These exchanges are represented by the short dialogue in (12). Speakers’ responses to the first and second interruptions were not scripted however, so that what they actually said, apart from repeating the target word itself, varied from token to token. The following is a typical exchange. 9 In fact, one such word was mistakenly included in the /o/ category as well: /bogatr/, [btr], ‘a hero of the Russian oral epic tradition’. Appearances notwithstanding, this word does not in fact share a root with /bog/, [bo], ‘god’, /bogatj/, [batj], ‘rich’, /ubogij/ [uboij], ‘wretched, destitute’, and so on. Rather, it seems to be a borrowing of early Turkic *baγatur, whence also Hungarian bátor, ‘brave’ [Vasmer, 1950-58]. Whether or not speakers of Modern Russian infer a connection between these forms is unclear. More appropriate for inclusion would have been, for example, bogatet’ [btet], ‘get rich’. This oversight did not affect results in any case. Sample dialogue (12) Speaker: Experimenter: Speaker: Experimenter: Speaker: Po-moemu on X skazal. Aaa? Ne ponjal. X on skazal! Što eto on skazal??!! X! ‘It seems to me he said X’ ‘Huh? I didn't get that.’ ‘He said X!’ ‘He said what??!!’ ‘X!’ Each target word thus received three repetitions, the second and third generally with substantial hyperarticulation. 3.1.3 Recording and analysis Sessions were recorded in a sound-attenuated room at Boston University at a sampling rate of 44.1 KHz, using a Marantz CDR-300 portable CD-recorder and a Shure SM10A cardioid headworn dynamic microphone run through a Rolls MP13 Mini-Mic preamplifier. Recordings were downsampled to 22.050 KHz and saved as .wav files using Praat 4.0.2 speech analysis software [Boersma and Weenink, 2004]. Measures of duration for target vowels were taken from linked spectrogram and waveform displays. After stops, vowels were marked as beginning at the onset of reliable periodicity following the burst. (Russian voiceless stops typically have low voice onset times.) Following laterals, vowels were measured as beginning at visible spectral discontinuities associated with the lateral’s release, or failing this, at the onset of significant energy in the second formant and above. In certain extreme cases, where no sufficient landmark could be detected between the lateral and the vowel, tokens were simply discarded. Measurements of the first formant, taken to be indicative of vowel height, and hence position on a potential cline of values between [a] and [], were made at vowels’ midpoints using LPC analysis (10 poles, with downsampling of original files to 11 KHz for female subjects and 10 KHz for male, analysis window 25 ms., shortened to 10 ms. where vowel durations necessitated). After discarding all tokens that could not be reliably measured (whether due to devoicing, overlap by neighboring consonants, or lack of sufficient landmarks to properly segment), of 216 total productions (18 words x 3 repetitions x 4 speakers), the total number of tokens analyzed was 209 (51 from RR, 53 from TM, 52 from DT, and 53 from M). Of these, 103 were instances of /a/ in Degree 2 reduction, and 106 were instances of /o/ in Degree 2 reduction. 3.2 3.2.1 Results Neutralization This experiment was not designed to be a definitive test of the completeness of neutralization of Russian Degree 2 reduction of /a/ and /o/. Had it been, it would have been necessary to control, for example, identity of surrounding consonants in a way that was unnecessary for present purposes. Additionally, the variance in both duration and F1 introduced by the manipulation of level of hyperarticulation in responses makes the comparison of category means for different underlying vowels difficult, in that this variability could easily mask the more subtle variation introduced by underlying vowel as a factor. As the question of phonological neutralization is central to the analysis of VR in Russian presented here, however, it must be noted that the results of this study do in fact support the conclusions of previous experimental work, particularly Padgett and Tabain [2005], on this question. As mentioned above, Padgett and Tabain find some (rather startling) evidence that certain reduction processes traditionally described as neutralizing are in fact not, with minute differences of the type familiar from the literature on incomplete neutralization persisting between mean realizations of seemingly merged underlying categories. The neutralization of /a/ and /o/ in their data, however, was found to be complete by any measure they applied. In the present study, results come only from the comparison of /a/ and /o/ in terms of duration and F1. Measures of higher formants were taken, but because no attempt was made to control place of articulation of surrounding consonants, and because, as noted above, different consonants affect undershoot of F2 targets in crucially different ways, error in the F2 data from the present study was too great to say anything of value on the question. In Table 3 and Table 4, F1 and duration means and standard deviations are shown for both vowels from each speaker. Note that the high standard deviations here reflect the intent of the experiment to produce each word in both ordinary and hyperarticulated variants. These values are thus effectively a measure of the greater success of the hyperarticulation task in producing a range of duration and F1 values than was achieved in the speech rate study of Barnes [2006], where standard deviations (see section 2.5 above) were much lower. [Table 3 about here] [Table 4 about here] Repeated measures analysis of variance performed on the data with underlying vowel as a within-subjects factor found no significant effect of underlying vowel on first formant measurements: F(1,3) = 8.509, p > .05. No main effect was found on duration either: F(1,3) = 7.207, p > .05. This is perhaps unsurprising, of course, given the low power of the test and degree of variability in the data. As a check on this result, individual univariate ANOVAs were carried out for each speaker assessing the effect of underlying vowel on both F1 and duration. Ordinarily of course, significance tests of this sort, however common they may be in the literature, are to be avoided; treating in this fashion individual trials from a single subject as independent experimental units is problematic in a variety of ways, most notably because it violates the assumption of independence of error effects, as well as massively overestimating the degrees of freedom for the error term in the analysis [Max and Onghena, 1999]. To the extent that the effect of these violations is an increased danger of Type I errors, though, swe can feel comfortable with the result that even under such circumstances, in the study at hand F1 and durational differences between /a/ and /o/ failed to reach significance for any speaker. These results are given in Table 5. [Table 5 about here] Still, it is interesting to note in Table 3 the tendency for each speaker’s underlying /a/ to have a higher mean first formant value than that of underlying /o/, with mean F1 for /a/ across all speakers at 576 Hz, compared with that /o/ at 552. Since stressed [a] has an F1 that is significantly higher than stressed [o] (see section 2.5 above), this difference in the present study could conceivably be taken as evidence of incomplete neutralization of /a/ and /o/ in unstressed syllables, a tiny lingering vowel height difference between what otherwise appear to be identical sets of vowels. Given the lack of statistical significance of this result, along with previous results in the literature, it is unlikely that this is the case. Indeed, one possible source for the oddly consistent F1 difference in this data is the fact that while all instances of underlying /a/ in this study are located in lexical root morphemes, several of the instances of underlying /o/ were in fact contained in prefixes, specifically /do-/, /po-/ and /pod-/. (See Appendix for a complete word list.) This is relevant due to a much-cited tendency in the literature [e.g., Bybee, 2001; Fidelholtz, 1975; Hooper, 1976; Jurafsky, et al., 2001; Zipf 1929] for more frequent or predictable (more probable, in Jurafsky et al.’s terms) material to be subject to more severe reduction processes than less frequent or predictable material. Grammatical morphemes such as verbal prefixes in Slavic languages are a prime example of such material. They are extremely (type-) frequent morphemes in that most verb roots in the language can appear with one or (usually) more such prefixes, and they are highly predictable in that they are members of a small closed-class of morphemes, not all of which commonly occur with every verb root. And indeed, when the underlying /o/ data from this study is broken down further, into those forms where /o/ is contained in a root (N = 5) and those in which /o/ is contained in a prefix (N = 4), the predicted difference emerges, as seen in Table 6. [Table 6 about here] Root /o/ is on average both longer and (by the analysis presented here hence) lower than prefix /o/. Using only the data from root /o/, mean F1 for underlying /o/ and /a/ become substantially closer together, if not identical. There is also one additional source of evidence for the completeness of the merger of underlying /a/ and /o/ in unstressed syllables in Russian, and that is their behavior under hyperarticulation. This issue is addressed more completely below in sections 4.5 and 4.6. For now, though, let us assume just the following: hyperarticulation is a device that serves at least in part to enhance phonological contrasts (witness the expanded vowel spaces discussed above from the literature on Clear Speech). Thus, if there is a contrast of any kind left between underlying mid and low vowels in Russian unstressed syllables, we might expect hyperarticulation to exaggerate that contrast, making it more readily observable.10 Phonological /a/, in other words, should become more [a]-like under hyperarticulation, while phonological /o/ ought to become more [o]-like. (The debt this line of reasoning owes to Gussenhoven and Rietveld [2000] is acknowledged more explicitly below.) Figure 3 shows the positions of stressed /a/ and /o/ (mean values taken from Barnes [2006]) as landmarks in a two-dimensional F1/F2 vowel space, and then plots relative to those reference points the mean formant values for the first, second, and third repetitions of unstressed underlying /a/ and /o/ for all speakers in the hyperarticulation study. Recall that the first repetition reflects something like ordinary speech, while the second and third repetitions should have exhibited progressively increasing hyperarticulation. This seems indeed to have been the case. [Figure 3 about here] Figure 3 shows clearly that both underlying /a/, and more importantly also underlying /o/ get progressively lower and fronter in their realizations as hyperarticulation increases. They both become, in other words, more [a]-like, in terms of both F1 and F2. Were there a surface contrast left to enhance between /a/ and /o/ in unstressed positions in this data, we would expect hyperarticulated tokens of underlying /o/ to move instead higher and backer in their realization, 10 There may seem to be some question-begging in the offing here concerning the proper analysis of incomplete neutralization. Competing points of view on this matter are addressed more completely in section 4.5. I would argue, however, most if not all of these seem to predict the same thing with respect to hyperarticulation, however they may differ otherwise. the very opposite of what in fact occurred. I take this to be strong evidence for the completeness of the neutralization of unstressed /a/ and /o/ in Russian. 3.2.2 Vowel Height as a Function of Vowel Duration The scatter plots in Figure 4 show first formant values in Hertz for underlying /a/ and /o/ for each speaker given as a function of vowel duration in milliseconds. Table 7 shows a linear regression model of the data for each speaker together with its R2 value. The differences in R2 values reflect the presence of a certain number of strongly influential cases in the data from some of the subjects (e.g., TM and M), on which more below. For now, it can be observed merely that extreme hyperarticulation (that is, realization of unstressed /a/ or /o/ with duration and F1 approaching values typical of stressed [a]) was relatively rare as a response, though it did occur. [Figure 4 about here] Furthermore, some speakers (again, TM and M) were more likely to produce responses with extreme hyperarticulation than others. This seems essentially to have been a consequence of the strategy adopted by each speaker for dealing with their interlocutor’s persistent requests for repetition: Some speakers, such as TM, became (or rather, I sincerely hope, pretended to become) visibly and audibly exasperated, with each repetition increasing in duration, loudness, and pitch range. Other speakers, such as RR, took a more stoic approach, quietly repeating the target word in each case without strong overt signals of impatience or annoyance. The difference is made graphically clear in Figure 5 and Figure 6. For stoic speakers, the hyperarticulation paradigm was simply less successful in eliciting extreme values for the relevant parameters than it was for other speakers—witness the comparatively much lower durations for all responses by RR. Indeed, the fact that even for the demonstrative TM more extreme duration and F1 values appear to create unacceptable leverage in the regression model, rather than an even cline of values from less to more hyperarticulated indicates that the hyperarticulation paradigm as implemented was less successful than we might have hoped. See section 4 for further discussion. What is clear from Figure 4 in any case, though, is that for all speakers tested, vowel duration emerges as a strong predictor of vowel height for the reduced vowels produced in this study. [Figure 5 about here] [Figure 6 about here] Figure 7 shows the data for all speakers pooled, 209 responses in total. Regression coefficients β0 (455.98) and β1 (2.3379) were both significant at less than .001 (t = 50.944 and 14.91 respectively). The overall model was significant at less than .001 (F = 222.304), with an R2 value of 0.5178. [Figure 7 about here] While clearly an important predictor of vowel height, the real significance of vowel duration in determining the extent of Degree 2 reduction is nonetheless masked to some extent in the data above due to the difficulty in comparing formant values across speakers, and in particular, across speakers of different genders, as is the case here. To remedy this situation, standardized scores for duration and F1 were produced individually for each speaker, and then these too were analyzed using linear regression.11 The result is shown in Figure 8. This amendment to the analysis has the effect of raising R2 for the model to 0.6376, a very high level indeed, indicating that nearly two-thirds of the variance in the F1 scores is explained by duration alone. I take this as strong support for the hypothesis that Degree 2 vowel reduction in Russian is a phonetic process, whereby underlying /a/ and /o/ are raised gradiently in the vowel space toward schwa as a function of durational compression. In cases of extreme hyperarticulation, both underlying vowels are produced with duration and F1 values similar to those ordinarily found in Russian stressed /a/. At the other end of the scale, when both vowels are extremely short, they are produced with a schwa-like quality. This relationship is mischaracterized when Degree 2 reduction is taken to be a phonological process whereby the symbolic category labels /a/ and /o/ are replaced by a category label // in the relevant unstressed syllables. Rather, only a single non-high vowel category, /a/, is licensed in any unstressed syllable, with F1 targets for the implementation of that vowel being determined to a large extent by the duration allotted to the unstressed syllable in question. [Figure 8 about here] Looking more closely at the regression model presented here, though, outlier cases of two types nonetheless emerge and demand an accounting. True outliers, that is, cases that deviate from the values predicted by the regression model were relatively few here. Examining the standardized residuals from the model, we find 10 cases at more than 2 standard deviations from predicted values. This is just slightly less than 5% of the total, and thus acceptable assuming that the residuals are normally distributed. Slightly more worrisome is the number of cases exceeding 2.5 standard deviations from the model’s predictions. There are 6 such cases, where a normal distribution of residuals would predict something more like 2 (i.e. approximately 1%). Interestingly, of the 10 cases falling farther than 2 standard deviations from predicted values, only one had a residual that was negative. All the rest were positive, meaning that they were instances in which F1 was higher (i.e. the vowel was less reduced) than its duration would predict. What is more, 8 out of these 9 “underreduced” vowels were produced by a single speaker, DT, in either his second or his third repetitions of a given target word. In each case, where our model would predict a lowish, schwa-like F1 (somewhere between 500 and 600), DT nonetheless realized his unstressed vowel with a relatively high, [a]-like value for F1 (700 or above). Since these are all second and third repetitions, another way of viewing these outliers is that DT, unlike the other speakers, seems able to hyperarticulate vowels by some other means than increasing their duration. This pattern is strongly reminiscent of a similar one discussed above in section 2.5 with respect to the results of the speech rate study of Barnes [2006]. Recall that there too the pattern of undershoot found in the productions of Speaker DT appeared to be difficult to explain with duration alone. Specifically, DT’s first pretonic vowels at short durations appeared to be systematically less reduced than the duration-dependent undershoot hypothesis would have predicted. There it was hypothesized, after Moon and Lindblom’s [1994] analysis of Clear Speech in English, that DT was avoiding undershoot at short durations by increasing his level of articulatory effort (or more concretely, his gestural velocities) in the production of the relevant syllables. While the data is admittedly sparse for drawing firm conclusions, in principle at least the same account could be extended to DT’s outlier cases in the present study. In effect, 11 Z-scores were used for standardization, calculated as zx = ( x - mx ) / sx, , where m is the mean for a given set of values, and s the standard deviation. our regression model for DT is actually missing a parameter: Rather than increasing segment durations to avoid undershoot in hyperarticulation contexts, DT is again adjusting his degree of articulatory effort instead. If this is true, and DT therefore represents a distinct subpopulation from our other speakers with respect to patterns of hyperarticulation, we might take this as justification for excluding his responses altogether from the model constructed here. Note that doing so improves the fit of the model substantially, as can be seen in Figure 9. In this new model, R2 becomes 0.678, and when again applied to standardized duration and F1 values instead of raw scores, R2 rises all the way to 0.707. [Figure 9 about here] There is, however, still the matter of the influential cases mentioned above, those cases that, though they do lie close to the line of best fit, nonetheless have values for the predictor variable extreme enough potentially to have an undue effect on the course of that line. The removal of such cases would alter both the regression coefficients and the degree of the model’s fit. Whether such cases belong in the model or not (i.e. whether their influence is indeed undue) is open to interpretation. I would argue that they do indeed belong in the model, that their extreme values for duration represent simply the far end of the hyperarticulation scale, yielding extreme, [a]-like values for F1, just as the analysis here predicts they should. Their relative rarity and distance from the bulk of responses stems simply from the failure of the hyperarticulation paradigm as implemented here to produce enough sufficiently hyperarticulated forms for analysis. Speakers, in other words, were generally too reserved in their responses to bring out the full range of values that might have been observed with a more effective elicitation strategy (on which problem, see section 4). To play devil’s advocate for a moment, however, we might conclude instead that these influential cases are skewing our understanding of the relationship between duration and vowel height in Russian unstressed syllables, and therefore deserve to be excluded from our analysis. To use a common diagnostic for influential cases, we could target for exclusion any responses with leverage values greater than three times the average (i.e. 3(k+1)/n, where k is the number of predictors, here one, and n is the number of cases included in the model). For our data set, the limit for the leverage statistic (as calculated by SPSS) would be .027. Applying this guideline, we locate 9 suspect cases. Interestingly, all but one of these are second or third repetitions of a target word by a single speaker, TM. Recall from above that TM was the most given to hyperarticulation of the four speakers, making some of her productions stand out from the general mean. Exclusion of these cases does change matters. R2 now falls to .302 for the raw values, or .507 for the standardized scores.12 Even so, however, the model still accounts for just over half of the variance in F1 scores, and is significant at p < .001. A scatter plot of the data with these cases removed is shown in Figure 10. [Figure 10 about here] 12 Adopting a more stringent view of what constitutes an influential case, setting the leverage statistic cutoff at just twice average (here .018), removes 13 cases, again all but one from TM's second and third repetitions. R2 is correspondingly reduced to .27 for the raw values, and to .495 for the standardized scores. 4 4.1 Discussion Vowel Duration in the hyperarticulation task Since the hyperarticulation experiment reported here was originally undertaken primarily to improve on the results of the speech rate study of Barnes [2006] in terms of the range of duration and F1 values obtained for /a/ and /o/ in circumstances of Degree 2 reduction, the first question to be decided in evaluating this study is whether or not it succeeded in this goal. Comparing the results of the two studies, it is clear that it did, though perhaps not to the extent that we might have hoped. Figure 11 illustrates the elevated duration and F1 ranges obtained in this study, as compared with those found in the speech rate study of Barnes [2006]. For all speakers, the hyperarticulation task has elicited higher values on both measures, as evidenced by the median, interquartile ranges, and overall ranges. The fact that the hyperarticulation task has not elicited duration and F1 values as low as were found in the speech rate task owes to the fact that in that study speakers were asked to read half the repetitions of each target sentence as fast as they could manage. In the hyperarticulation task, on the other hand, the lowest durations and F1 values would have occurred in the non-hyperarticulated first repetitions of each target sentence, and these would have been produced at something like a neutral, conversational rate of speech. Clear also from Figure 11 is the fact, noted already above, that one speaker, TM, produced hyperarticulated forms with more abandon than the remaining speakers. RR in particular altered her productions little as a result of the experimenter's prompts for repetition. This difference in response strategy is reminiscent of the findings of Curl [2002], to the effect that the extent to which requests for repetition will in fact elicit “clear speech” is dependent on the pragmatics of the speech situation, and more specifically, on the extent to which the speaker finds the request for repetition “justified”. In this light, RR seems to have responded as though the experimenter’s requests for repetition were in fact warranted by some aspect of her source productions, while TM responded as though the requests were unwarranted. Be this as it may, the range of responses to the task suggests that another means of eliciting clear speech might have captured a fuller range of duration and F1 values for all speakers, and thus would be preferable for future research in this direction. More hyperarticulation from the remaining three speakers, for example, would have made TM's productions seem less extreme in the pooled analysis. One common procedure that might be an improvement would be to ask speakers to read as though their listener had a hearing impairment, or less-than-fluent knowledge of the language (as in, e.g., Smiljanić and Bradlow [2005]). [Figure 11 about here] 4.2 Implications for the analysis of Russian vowel reduction Such minor defects notwithstanding, the results of the hyperarticulation study lend strong support to the conclusions of Barnes [2006] concerning the nature of Russian vowel reduction: for the Degree 2 reduction pattern investigated here, all speakers showed a highly significant correlation between vowel duration and vowel height as measured by F1. Contrary to all previous reports, reduction of /a, o/ to [] in fact applies gradiently, with significant raising in shorter vowels, and substantially less in syllables made longer through hyperarticulation; where durational values of second pretonic vowels approached or coincided with values characteristic of first pretonic or stressed vowels, F1 increased analogously. Indeed, in the most hyperarticulated tokens of second pretonic vowels, both duration and F1 were well within the range normally characteristic of /a/ bearing lexical stress. Additionally, these hyperarticulated phonetic [a] vowels were found equally realizing underlying phonological /a/ and of underlying phonological /o/. Which is to say, while increased duration was sufficient to allow realization of an unraised phonetic [a] in a lexically unstressed syllable, no amount of lengthening was sufficient to realize phonetic [o] in that same environment. That out of 209 productions not a single speaker, under any degree of hyperarticulation, produced so much as a single unstressed phonetic [o] argues forcefully for the conclusion that the neutralization of Russian /a/ and /o/ in unstressed syllables is categorical and phonological, displaying no sensitivity at all to the phonetic context in which the vowels are produced on any given occasion. It also argues against an interpretation of the hyperarticulation task as somehow paralinguistic or unnatural. Were hyperarticulation indeed something more like spelling pronunciation, for example, we might expect to hear underlying /o/ surface intact when hyperarticulated, since Russian orthography maintains a distinction between ‘a’ and ‘o’ in spelling even where that contrast is neutralized phonologically. Based on these findings, I argue that we must revise the traditional analysis of Russian vowel reduction as follows: there are indeed two types of vowel reduction corresponding in a loose way to what have been called Degree 1 and Degree 2 of reduction. The difference between the two patterns, however, is not one of degree, or even of functional motivation à la Crosswhite, but rather of the level of representation at which the two patterns are realized. To the extent that gradient application is a hallmark of phonetic processes, we may conclude that Russian Degree 2 reduction is a matter of phonetic implementation, the result of duration-dependent undershoot of F1 targets in durationally impoverished unstressed vowels. The best model of Russian /a, o/ → [] will thus be something like the decaying exponential functions used by Moon [1991] and Moon and Lindblom [1994] (after Lindblom [1963]) to model the English and Swedish data in their respective experiments, an equation of the general form F1 = –αe-βt + F1target. (This is a simplification in some respects of the equations offered by Lindblom and Moon, whose control of consonantal context allowed them to build consonant place into their model along with vowel target and duration. It does express at least the desired shape of the function, rising toward an asymptote [the F1 target] as t [time] increases, and falling exponentially as t decreases). Such an equation should describe the realization of /a/ in any structural context in Russian, stressed or unstressed. That a schwa-like vowel surfaces in the second pretonic syllable, but does not in the first pretonic and stressed syllables, is not the result of the application of two distinct reduction processes, but rather just a consequence of the different amounts of duration normally associated with each of those syllables. As durational values change, whether due to speech rate, or to hyperarticulation, vowel height changes too. Neutralization of /a/ and /o/, on the other hand, is a phonological process, and as such it is best modeled, for example, as the result of an Optimality-theoretic constraint banning mid vowels in unstressed syllables (or licensing them only in stressed syllables), such as the following. (13) License (–high, –low) / σ This constraint makes no mention of phonetic duration, being defined solely in terms of elements of categorical phonological structure. As such, it will apply in equal measure to any string meeting its structural description, regardless of factors from phonetic context. (The interaction of this constraint with relevant Faithfulness constraints will determine what in fact surfaces in unstressed syllables that contain underlying mid vowels, i.e. whether alternations will involve raising or lowering. See Beckman [1998], Crosswhite [2001], and Zoll [1998] for various models of alternations such as these in Optimality Theory. The details of these accounts are beyond our concerns here.) The result of accomplishing raising to schwa in the phonetics is that Russian phonology no longer needs to license a representation // in any position, stressed or unstressed, and hence also escapes the need to refer in any way to the unnatural class of positions in which Degree 2 reduction fails to occur. Non-reduction to schwa in first pretonic syllables, absolute word-initial position, and absolute phrase-final position are all a consequence of the additional duration accorded to the phonologically low vowels realized in these positions. Characteristic high F1 for [+low] vowels in these positions is a consequence of the output of a function like that sketched above, rather than of what are effectively exception clauses to a phonological constraint that otherwise would create schwa. This analysis simplifies Russian phonology greatly, while requiring nothing additional from the phonetics, save the capacity to realize subphonemic durational patterns related to prosodic boundaries and accent in a way that is at once in line with well-documented cross-linguistic patterns, but also capable of significant variation in a languagespecific way. 4.3 Effort and undershoot: speaker- and position-specific patterns There are, nonetheless, ways in which the analysis presented in the preceding paragraphs oversimplifies matters. What can be concluded from the above is that for all four speakers, there was a correlation between duration and F1 for second pretonic /a, o/, such that vowels with greater durations approached, and in many cases overlapped with, expected F1 values for low vowels in first pretonic and stressed syllables, strongly suggesting that reduction to schwa in Russian is a gradient process subject to reversal in the correct phonetic contexts. Certain aspects of the results, however, raised questions as to whether duration by itself is sufficient to predict degree of undershoot in Russian. In the hyperarticulation study presented above, the productions of speaker DT in particular stood out as reminscent of earlier work on hyperarticulation in English by Lindblom [1990] and Moon and Lindblom [1994]. Recall that these studies argue for the inclusion of articulatory effort as a variable in a model of F2 undershoot in English. Like the “clear speech” productions in those earlier studies, a substantial number of DT’s hyperarticulated productions in this study stood out from the rest of the results as realizing surprisingly little undershoot for the amount of durational compression involved. The interpretation offered in the previous section for this is that DT is employing a strategy of increased effort, rather than (just) added duration, to avoid undershoot in his hyperarticulated speech. This type of interspeaker variation in the realization of clear speech would fall in line nicely with the results of other studies, such as Kuehn and Moll [1976], which suggest that different speakers may respond differently to durational pressure on articulation—some by yielding to undershoot, others by increasing articulator velocities to compensate. Whether DT is anomalous in this respect, or characteristic of a broader class of Russian speakers is impossible to say from a study with as small a sample of speakers as this one. Clearly though, this is an important question for future research to address. It was also noted above that DT stood out again in the comparison of first pretonic and second pretonic syllables in the speech rate study of Barnes [2006]. While some speakers, such as TM, seemed to produce a single smooth curve of F1 values spanning both prosodic contexts, DT's results seemed different; his second pretonic /a/ and /o/ did exhibit raised F1 as predicted at greater durations, but his first pretonic vowels seemed to maintain relatively high F1 values even in the face of substantially decreased vowel durations. Thus, for a given duration, DT's first pretonic F1 values were consistently higher than those of his analogous second pretonic vowels. Recalling de Jong's [1995] analysis of the articulatory profile of English lexically stressed syllables as localized implementation of Lindblom's hyperarticulatory mode, we may wonder whether, at least for some speakers of Russian, the first pretonic syllable constitutes just such a domain of localized hyperarticulation (hereafter DLH). Whether the first pretonic syllable constitutes a distinct DLH whose level of contrast enhancement must be defined separately from that of the stressed syllable, or whether for speakers such as DT the first pretonic and the stressed syllable are better analyzed as a single, extended DLH remains to be seen. We will also want to know how these DLH’s are like or unlike the other undershoot-resisting contexts in Russian, absolute word-initial and absolute phrase-final position. In particular, how, if at all, does the enhancement of clarity differ in its implementation from DLH to DLH in Russian? Likewise, how, if at all, does the type of contrast enhancement we find in Russian differ from that documented in other languages? Differences in either instance might be quantitative, in terms of degree of additional effort, or they could be qualitative, in terms of the phonetic features selected for enhancement. This research program is already well underway in the contexts of boundaryadjacent and accentual enhancement patterns in a number of languages—see the references in section 2.5. In the Russian context though, what is most urgently needed is a follow-up study of the variables contributing to undershoot. This study would involve a larger sample of speakers than did this one, and it would restrict participants to native speakers of a single regional variant of Contemporary Standard Russian (i.e. all Moscow speakers, or all Petersburg speakers, on which see the following section). Secondly, target contexts for comparison would include not only second and first pretonic syllables, but also the other potential DLH’s discussed above, such as absolute word initial position and absolute phrase-final position. Each of these positions should be produced in a range of contexts, including extreme durational compression (fast speech), ordinary or conversational speech, and hyperarticulation. It would then be possible to determine whether a single function suffices to describe the relationship between duration and F1 in all these contexts, or whether susceptibility to undershoot differs from context to context. 4.4 On vowel reduction and causation in phonology The results of this study also bear interestingly on debate concerning the nature of the causal connection between duration-dependent undershoot and phonological vowel reduction. Recall from section 2.4 that such a causal connection, if we accept it, could be viewed as active in at least two distinct ways: Barnes [2002, 2006] proposes a diachronic link between the two phenomena; severe duration-dependent undershoot, on this view, is the historical precursor to systems of phonological vowel reduction, which arise out of the former when perceptually compromised contrasts are lost through sound change. One crucial prediction of this approach is that once phonologized, patterns of neutralization should no longer be dependent on the facts of phonetic duration for their continued existence; the patterns become abstract, structural generalizations about the distribution of vowels in the language, rather than online adaptations to the vagaries of phonetic context. The results of the present study are consistent with this view in the following ways: raising of /a/ to schwa, as a phonetic process, is subject to variation as a response to changes in the facts of phonetic context in any given utterance. Durational changes resulting from changes in speech rate or degree of hyperarticulation will cause proportional changes in the extent of application of the raising process, and this is precisely what the present study finds. The neutralization of /a/ and /o/ in unstressed syllables, on the other hand, is a phonological pattern. As such it is not sensitive to the phonetic characteristics of the context in which any given form is uttered; increases in duration resulting from a slower rate of speech or from hyperarticulation do nothing to counteract the neutralization of /a/ and /o/ because this neutralization depends not on the phonetic durations found in lexically unstressed syllables, but to the very unstressedness of those syllables itself, as an abstract generalization about the metrical structure of the prosodic word in Russian. No amount of additional duration can change that metrical structure, because syllables are not stressed or unstressed in Russian by virtue of the durations they bear, but rather the reverse. This prediction too is upheld by the present study; while hyperarticulation often produced formant values for unstressed underlying /a/ and /o/ well within the range typical for an /a/ bearing lexical stress (“undoing” the effects of Degree 2 reduction), no amount of hyperarticulation was able to undo the neutralization of the contrast between /a/ and /o/; nothing like an [o] ever surfaced in an unstressed syllable in this study.13 An opposing view of the relationship between duration-dependent undershoot and phonological vowel reduction holds the causal connection between the two phenomena to be active synchronically, a fact about the grammar of any language exhibiting phonological vowel reduction [Flemming, 1995/2002, 2001, 2006]. Phonological vowel reduction arises when the grammar of a language mandates that speakers give up certain contrasts to neutralization rather than expending the effort to realize them in circumstances that would compromise them perceptually. Significant duration-dependent undershoot is thus a necessary condition for the manifestation of phonological vowel reduction (though it is perhaps not a sufficient condition, in that the grammar of a language with undershoot but no reduction can always just be analyzed as more permissive of perceptual difficulty than a language that exhibits neutralization of contrasts under similar circumstances). The facts of Russian challenge this view in several ways. The first problem is well-known from the literature on phonetically-driven phonology, and is exemplified by the results of the hyperarticulation study: if phonological reduction is really a consequence of duration-dependent undershoot, we might expect that the additional duration contributed to unstressed vowels under hyperarticulation would be sufficient to license the reappearance of contrasts such as that between Russian unstressed /a/ and /o/. That this does not occur is an instance of what Hayes and Steriade [2003] call the stabilization problem. The idea is that the phonology must somehow achieve a relatively stable form despite massive variation in phonetic realization across contexts. 13 It should be noted that the diachrony behind the phonologization of this pattern is a matter of dispute in the Slavist literature. See Barnes [2006: 36-37] for details. For this to happen, some of that variation must be ignored. For speech rate and hyperarticulation to be as irrelevant as they are for the application of Russian phonological vowel reduction, any synchronic causal connection between duration-dependent undershoot and phonological neutralization must not be a direct one. Rather, it must be mediated in such a way that the phonology refers not to the actual duration of any given vowel in performance, but instead either to phonetic duration in some neutral or average performance context [Steriade, 1999], or to some average durational value computed across all possible contexts [Hayes, 1999]. Either way, the notion that sound patterns depend on “average” duration rather than duration itself makes the causality we understand to be involved here more abstract than might be supposed at first brush. Reduction to schwa in Russian, by contrast, seems to involve actual duration, and hence also actual causality. The second problem for the synchronic causality view comes from the durational and F1 values obtained in the first pretonic syllable in the speech rate study presented in Barnes [2006] and summarized here in section 2.5. What is troublesome here is that while mean duration values for first pretonic /a/ and /o/ were somewhat lower than the stressed vowels measured (approximately 70 ms as compared with 82 and 75 ms for stressed /a/ and /o/ respectively)14, mean F1 values for first pretonic /a/ and /o/ were not any lower than those obtained for stressed /a/ (705 and 694 Hz for first pretonic /a/ and /o/ respectively as compared with 700 Hz for stressed /a/). Interestingly, this remains true even when speakers’ “fast” productions are factored out and only "normal" or "conversational" speech rate productions are considered: mean F1 for conversational productions of first pretonic /a/ and /o/ was 717 Hz, as compared to 708 Hz for stressed /a/. What this all means is that the low vowel being realized in first pretonic syllables here is not in fact undershot at all in comparison with analogous stressed vowels. To the extent that these productions are also “average” or “typical”, an account of Russian phonological vowel reduction driven by synchronic duration-dependent undershoot encounters a problem here: the vowel space in first pretonic syllables is not in fact compressed, and lacking such compression, we would predict that no neutralization of contrasts would be necessary. And yet it happens anyway. It should be noted that this result is strange also in light of traditional phonetic descriptions of the realization of first pretonic /a/ and /o/ in Russian. As noted above, this vowel is usually transcribed '', a symbol from the Slavist tradition which corresponds to something like IPA [], a vowel somewhere in between low [a] and reduced, mid schwa. The traditional description is reflected to some extent in the instrumental results of Padgett and Tabain [2005], who document raising in first pretonic syllables to values somewhere in between those of the [a] of stressed syllables and the schwa of the second pretonic. Padgett and Tabain also note, however, that the mean amount of raising they obtained for first pretonic /a/ and /o/ seemed surprisingly small given traditional descriptions of the height of this vowel. Taking values from their Table 3 [2005: 30], the mean difference between F1 for low vowels in stressed and first pretonic positions (in non-palatalized contexts for all female speakers in their study [N = 8] was 1.1 ERB, or just above the 1 ERB value adopted by them as a rough estimate of the threshold of perceptibility for differences in formant values. They note that such a small 14 That stressed vowel durations were as low as they are here is a consequence of the manipulation of speech rate in this study: recall that half of each subject’s productions came as a result of an instruction to speak “as fast as possible”. difference is a potential problem for a theory of vowel reduction based on perceptual difficulties arising from duration-dependent undershoot. Could such a small degree of undershoot really compromise the distinctions between high and mid or mid and low vowels in the system? They suggest as a hypothesis, citing Crosswhite [2001] that a better measure of perceptual confusability might include some sort of weighting of formant differences for vowel duration, the idea being that the same spectral distance might be harder to distinguish on shorter vowels than on longer. Still, the fact that in both their study and that of Barnes [2006] first pretonic /a/ and /o/ showed typical durations in the vicinity of 70 to 90 milliseconds should make us wonder just how compromising such durations could realistically be of vowel contrasts, particularly in the absence of significant undershoot. Durations of this magnitude, after all, are more characteristic of the unstressed vowels of languages like Serbo-Croatian, Macedonian, Polish, or Czech (see above) that lack phonological vowel reduction, rather than of the much shorter unstressed (non-first pretonic) vowels of canonical phonological vowel reduction languages like Belarusian or Russian. (See above, section 2.4.) If durational values this large are truly compromising of vowel quality contrasts, then our typology of unstressed vowel reduction should look substantially different than it actually does. But this is not the whole story, for like Barnes [2006] and the present study, Padgett and Tabain failed to obtain a sample of speakers that was homogenous with respect to region of origin in the Russian-speaking world. Specifically, of the nine speakers included in their study, only four were born and raised in Russia itself, and only two of these were natives of Moscow. Three were actually born and raised in China, the descendents of White Guardists who fled Russia in the wake of the civil war from which the Soviet Union eventually emerged. This is not necessarily a problem of course; all participants in all three studies were carefully vetted as active native speakers of Contemporary Standard Russian, devoid of any obvious regional peculiarities. Even within the bounds of what is uncontroversially CSR, however, minor regional differences exist, and could have escaped detection. Most important for our purposes are differences in the pronunciation of unstressed /a/ and /o/ in the first pretonic syllable, and here a clear regional difference is in fact well-documented. The relevant division is described as being between Muscovites and non-Muscovites. Timberlake [2004: 81] describes the so-called ‘Old Muscovite’ pronunciation (i.e. the dialect of pre-1917 Moscow), one of the characteristic features of which was a substantially lower realization of first pretonic /a/ and /o/. Timberlake gives the standard transcription [α] for this vowel. Most features of Old Muscovite pronunciation, Timberlake notes, have since been overtaken by the broader national norms of CSR. This one, however, apparently has not. Indeed, an unusually long and low first pretonic [a] is now a widely recognized sociolinguistic marker of Muscovite origins.15 Looking again at Padgett and Tabain's data (in particular Table A4 in their appendix), we find confirmation of this peculiarity of Muscovite pronunciation. Comparing measures of vowel height for stressed /a/ and first pretonic merged /a/ and /o/ (given as F1–F0 measured in ERB), we see that all speakers show at least some undershoot in first pretonic syllables. For seven of 15 Purely anecdotally, ask nearly any native speaker of Russian for a characterization of a “Moscow accent”, and the result will usually be an enthusiastically cartoonish rendition of the phrase Ja iz Moskvy, ‘I am from Moscow’ (ordinarily [ja iz mskv]), rendered with first pretonic [a] in the initial syllable of Moscow extremely long, unmistakably low, and bearing a significant rising pitch contour. Dialect stereotypes are of course not the greatest source for phonetic data all things being equal, but nor do they arise altogether out of thin air. nine of speakers, these differences are 1 ERB or greater, meaning again that they are above what Padgett and Tabain use as a rough threshold of perceptual discriminability. (In fact, their speaker JD has first pretonic /a/ and /o/ .94 and .96 ERB below stressed /a/, and speaker TO has a first pretonic /o/ .93 ERB below stressed /a/. I take these differences to be close enough to the standard of 1 ERB, given the roughness of the boundary to begin with.) Two speakers in this study, however, the only two born and raised in Moscow, have first pretonic /a/ and /o/ that are significantly under this 1 ERB threshold away from stressed /a/: their speaker AC has F1–F0 for first pretonic /a/ and /o/ only .62 and .66 ERB lower than stressed /a/, while their speaker MK has first pretonic /a/ just .36 ERB below stressed /a/, while her first pretonic /o/ is actually .01 ERB above her stressed /a/. In sum, while most speakers measured by Padgett and Tabain did have some (though not much) raising of the low vowel in first pretonic position relative to its stressed syllable analogue, their two Muscovite speakers differed from the rest in showing so little raising of this vowel as to fall below the threshold of perceptibility set for vowel quality differences. If most speakers, in other words, showed surprisingly little undershoot in first pretonic syllables, the Moscow natives showed next to none. Clearly what is necessary is a follow-up study to Barnes [2006] and/or Padgett and Tabain [2005] in which speakers’ regional origins are controlled, or perhaps even set as a betweensubjects factor for investigation, comparing either Moscow vs. Non-Moscow, or perhaps better Moscow vs. some second location, such as St. Petersburg. The added difficulty in finding suitable speakers would doubtless be repaid by the increased accuracy of the picture we would obtain of vowel reduction in Russian. For the time being, at least, I will conclude here that the effective lack of F1 undershoot in first pretonic syllables of Modern Muscovite Russian is a serious problem for theories of vowel reduction that rely on duration-dependent undershoot as a necessary synchronic causal factor for phonological neutralization. No undershoot, not even in “average” or “neutral” contexts, should mean no neutralization, and yet neutralization occurs nonetheless. As mentioned above, this same problem is found also in the unstressed syllables of Belarusian (especially the first pretonic) and Shimakonde. (See Barnes [2006], Chapter 2 for references and details on both.) 4.5 On the enhancement of contrasts by hyperarticulation As reviewed briefly above, most of the work to date on clear speech or hyperarticulation has focused on (1) establishing the set of phonetic features that undergo modifications in this register in English, or (2) establishing the extent to which use of this register in English enhances intelligibility under a variety of circumstances in which communication is impeded (due either to characteristics of the listener or characteristics of the environment). As Smiljanić and Bradlow [2005] note, there has been almost no work on clear speech for the enhancement of intelligibility in languages other than English (their comparative study of English and Croatian being a notable exception). The present study thus adds to our general understanding of the nature of hyperarticulation in a small way by confirming that in Russian, as in English and Croatian, clear speech involves both extended segment durations and an expanded vowel space. (To be fair, this latter is actually somewhat more than we can say with confidence, since only low vowels were analyzed. At the very least, clear speech in Russian involves a lessening of the undershoot to which low vowels in unstressed syllables are otherwise subject.) More significantly though, this study also adds something to our understanding of clear speech which has not been demonstrated before. This is that the contrasts that are perceptually enhanced by hyperarticulation in Russian appear to be surface contrasts only. Contrasts at the level of Underlying Representation (or to put it another way, category membership at the level of the morphophonemic alternation set rather than category membership at the level of the individual phonological segment string) appear inaccessible to the activity of hyperarticulation. The evidence that the present study provides for this shallowness of contrast enhancement comes from the continued complete neutralization of underlying /a/ and /o/ in unstressed syllables, even under conditions of extreme hyperarticulation. By way of clarification, recall that under the assumptions laid out above, the neutralization of /a/ and /o/ in Russian is a process of phonological category relabelling, with the result that segments that were representationally distinct in the Input to the phonology become representationally identical in its Output. Imagining now for a moment a hypothetical set of circumstances under which hyperarticulatory contrast enhancement were not shallow, but rather deep, what we would expect to find would be modifications to the ordinary phonetic realization of a string such that the acoustic signal actually came to contain useful cues under hyperarticulation to contrasts at the level of Underlying Representation. Deep contrast enhancement would then constitute at least a partial “undoing” of phonological neutralization. We can also easily imagine why such a process would be functionally desirable: deep contrast enhancement would help to resolve surface ambiguities created by phonological neutralization. In the Russian case, deep enhancement would give us perceptible differences between underlying /a/ and underlying /o/ surfacing in unstressed syllables, with /o/ moving perhaps up and back in the vowel space, and /a/ moving analogously forward and down. Obviously, this is not what we observe (recalling Figure 3). Rather, extreme hyperarticulation of underlying /o/ resulted in production of something strongly reminiscent of stressed [a], precisely the result of extreme hyperarticulation of underlying /a/. The contrast enhancement we see here is thus shallow in the sense that it moves the realization of both underlying /a/ and /o/ toward ideal targets for a phonological /a/. It is the output of the phonological process of neutralization that is thus enhanced, rather than its input. One area in which the question of shallow rather than deep contrast enhancement becomes particularly interesting is in the context of certain recent work on incomplete neutralization. Incomplete neutralization is the phenomenon whereby underlyingly contrasting segments that appear to be neutralized on the surface (most famously final voiced and voiceless obstruents in languages such as Dutch, German, Catalan, and Polish) are nonetheless realized with subtle phonetic cues to their underlying identity [Dinnsen and Charles-Luce, 1984; Dinnsen and Garcia-Zamor, 1971; Fourakis and Iverson, 1984; Port and Crawford, 1989; Port and O’Dell, 1985; Slowiaczek and Dinnsen, 1985; Warner, et al., 2004]. In Dutch, for example, Warner et al. [2004] show that word-final underlying /t/ has a slightly longer burst duration on average than does word-final underlying /d/, and that vowels preceding word-final underlying /d/ are slightly longer on average than vowels preceding word-final underlying /t/, and all this despite that fact that word-final devoicing in Dutch is commonly held to neutralize the contrast between underlying voiced and voiceless obstruents. The problem this result potentially causes for what Pierrehumbert [2002] calls the traditional feed-forward modular view of the relationship between phonology and phonetics should be clear enough: if neutralization indeed takes place in the phonology, then it is achieved by the substitution of the phonological category label [–voice] for underlying [+voice] on obstruents in word-final position, with the result that underlying voiced and voiceless obstruents should exit the phonological module of the grammar with identical surface representations. To the extent that the input to the phonetic interpretation module is identical in both cases, the output of phonetic interpretation should on this view likewise be identical (assuming a strict view of informational encapsulation in the activity of distinct modules [Fodor 1983]). Scholars have differed wildly in their interpretations of the incomplete neutralization result over the years, with views ranging from rejection of the phenomenon as “linguistic” to begin with [Fourakis and Iverson 1984], to celebration of it as the clearest expression on record of the true nature of the relationship between phonetics and phonology [Gafos, 2006; Pierrehumbert, 2002; Port, passim]. Perhaps the most prominent current manifestation of this latter view [Ernestus and Baayen, 2006; Gafos, 2006; Port and Crawford, 1989; Yu, 2006] holds, in so many words, that incomplete neutralization is in fact nothing other than the deep enhancement of contrasts counterfactually imagined above in our hyperarticulation study. While researchers differ in the details of their proposals, held in common is the view that there is nothing unusual or “unphonological” about the phonological process responsible for final devoicing in languages displaying incomplete neutralization; indeed, neutralization is as complete or categorical in these cases as it is ever going to be. The appearance of “incompleteness” arises through a process I have elsewhere called morphophonetic gravitation, and which others have termed phonetic or subphonemic analogy [e.g., Steriade 2000]. Such a phenomenon, if real, would involve the biasing of production targets for a given segment in the direction of production targets for the corresponding segment in some other, morphologically related, surface form. Take again the example of the Dutch underlying final /d/ that, upon apparent phonological devoicing, nonetheless remains ever so slightly more /d/-like than instances of true /t/ in the same position. This “slight voicing” (as Ernestus and Baayen [2006] have termed it), according to the gravitation approach, is caused by the attractive force exerted by paradigmatically related forms in which the underlying /d/ is not in a position of surface neutralization, and hence surfaces voiced; a Dutch form such as Zweed, /zwed/~[zwet], ‘Swede’, comes to differ slightly from a form like Zweet, /zwet/~[zwet], ‘sweat’, through gravitation toward, or on analogy with, related forms such as Zweden, /zwedn/~[zwedn], ‘Sweden’. What is important here is the causal role played in all the works cited here by the morphophonemic alternant in the creation of these effectively skewed productions. The way in which morphophonetic gravitation resembles our deep enhancement should be clear: both involve the deployment of elements of subphonemic detail for the purpose of signaling lexical identity in spite of constraints on phonological representations that would otherwise obscure it. The question then becomes to what extent incomplete neutralization is in fact hyperarticulatory in nature. It could be, for example, that incomplete neutralization does involve biasing from the production targets of morphologically related forms, but that this biasing is neither voluntary nor gradable in the way that hyperarticulation typically is. In fact, such a view of gravitation as an involuntary consequence of the structure of the lexicon and the architecture of the speech production mechanism seems to be what is advocated by Pierrehumbert [2002], and is perhaps consistent with the account of phonetic analogy given by Steriade [2000] as well. Interestingly, though, several authors have identified precisely this kind of speaker-controlled, hyperarticulatory character as a hallmark of incomplete neutralization phenomena. The foundational result for this strain of thought is that of Port and Crawford [1989], who showed that the “incompleteness” of neutralization of voicing in German final obstruents differs in degree depending on the pragmatic context of the task used to elicit the forms for comparison. Specifically, voiced and voiceless final obstruents in Port and Crawford’s study showed the least differentiation from one another in situations most like ordinary speech (target words embedded in sentences read in non-contrastive contexts), and showed the most differentiation in the most metalinguistic, or least “natural” situation tested (dictation of isolation forms to a waiting transcriber). Gafos in particular focuses on this apparently voluntary, hyperarticulatory aspect of incomplete neutralization, as do Warner et al., who go further still in equating incomplete neutralization ultimately with phonetic differences arising from hypercorrect renderings of distinctions that are purely orthographic, rather than phonological in nature. Assuming that incomplete neutralization does have this sort of gradient, speaker-controlled implementation as a fundamental characteristic, the importance of the present study becomes immediately apparent. We have demonstrated that in the case of Russian vowel reduction, hyperarticulation results in the enhancement of phonological contrasts as expected, but that this enhancement is strictly shallow. The surface identity of underlying /a/ and /o/ appears impenetrable to hyperarticulation. Let us now assume that this is the case more generally, viz. that hyperarticulation is only ever capable of shallow enhancement, such that phonological neutralization is always opaque to it. If this is true, and if hyperarticulation does indeed enhance a contrast between word-final voiced and voiceless obstruents in Dutch (or German, or Polish...), then contrary to most analyses, that contrast must in fact be present (i.e. unneutralized) in the output of the phonology. But if this is true, then the phonological business-as-usual gravitation approach to incomplete neutralization must be wrong after all. In other words, there must indeed be something incomplete about incomplete neutralization. In Barnes [2006] I suggest the possibility that incomplete neutralizations are not in fact phonological neutralizations, but in fact just gradient processes of extreme contextual phonetic approximation of distinct phonological categories. This approach equates incomplete neutralization with the cases of near-merger familiar from the dialectological literature [Labov, Karan and Miller, 1991; Labov 1994] in a way that the gravitation approach has a harder time doing. There were, of course, several big “ifs” in the above proposal. It may be that some otherwise categorical neutralizations are in fact subject to Deep Enhancement under hyperarticulation, making Russian somehow anomalous in this respect. It might also turn out upon more comprehensive testing that clear speech does not in fact augment the incomplete neutralization result in the way that was initially reported. Both of these possibilities are real and need empirical testing in a larger number of cases before we can feel confident in our understanding of the situation. It may also turn out in the course of this testing that the distinction between a categorical phonology and a gradient phonetics is itself, as many have suggested, fundamentally ill-conceived, and as such inadequate to a proper analysis of the facts of these cases. The depth of contrast enhancement under hyperarticulation and phenomena such as incomplete neutralization will in any event I believe constitute central proving grounds for competing theories in this area. 4.6 On hyperarticulation as a tool for investigating category membership Beyond the questions raised in the preceding section, this study has shown that the elicitation of hyperarticulated or clear speech productions of target forms can be a valuable tool for assessing phonological category membership in cases where phonetic realization in ordinary speech leaves ambiguities of interpretation. As noted above, among the most consistent findings in the study of clear speech have been extended segment durations and an expanded vowel space, with at least a partially causal relationship between the former and the latter. Improved dispersion in the vowel space is believed to enhance intelligibility by increasing the perceptual distance between contrasting elements. The decrease in undershoot allows vowels to be realized closer to what we assume are their ideal targets, and hence, presumably, farther from one another, minimizing confusability. In some circumstances (e.g., minimal undershoot, no phonological neutralization), we have a clear sense of what this kind of contrast enhancement should mean phonetically, at least for peripheral vowels: front vowels should be fronter, and back vowels backer; high vowels should be higher, and low vowels lower. In other cases, however, such as where undershoot is more extreme, or phonological neutralizations have potentially taken place, it is less clear what we should expect from the enhancement of contrasts. The preceding section sketched competing scenarios for the latter problem under the rubrics of shallow and deep enhancement. To the extent that what we observe is in fact only shallow enhancement, the elicitation of hyperarticulated forms becomes a valuable methodological device for the determination of surface phonological category membership. The idea that hyperarticulation could actually be used as a diagnostic tool of this kind has so far, to my knowledge, been employed only in the study of intonational categories. (Note, however, the use of manipulation of speech rate [Solé 1992] or focus structure [de Jong 2004] to achieve similar ends in the area of contrasts in segmental phonology.) Acting on an assertion by Liberman and Pierrehumbert [1984], Gussenhoven and Rietveld [2000] investigate the distinction between two intonation contours in Dutch (“high rise” and “low rise”), the phonological identity of which had been a matter of uncertainty in the literature. One possibility is that the two contours are categorically distinct, the phonetic realization of contrasting phonological representations. The other possibility is that the distinction is gradient, with the two contours ultimately being variant realizations of a single phonological category. Gussenhoven and Rietveld propose to investigate the question through the manipulation of pitch span—effectively hyperarticulation. The premise of the experiment is that the phonological contrast in question, to the extent that one exists, is between a high pitch accent and a low pitch accent at the beginning of the string. Accordingly, if an accent is a phonological H, enhancement of contrasts under hyperarticulation should cause it to be realized higher than usual; if, on the other hand, it is actually an L, it should be realized lower under the same circumstances. If the relevant tones in the two contours move in opposite directions under hyperarticulation, the contours would be judged phonologically distinct, and the tones contrasting. The study Gussenhoven and Rietveld in fact carry out involves native-speaker evaluations of synthetic versions of the two pitch contours with pitch ranges expanded in the ways described above. Speakers were asked to rate the target contours for the degree of “surprise” they conveyed. (Since the relevant contours are for questions in Dutch, surprise seemed a reasonable way of describing emphasis or degree of involvement.) The result argued for a phonological contrast between the two contours: the lower the low rise contour was made to begin, the greater the degree of surprise listeners attributed to the utterance. Likewise, the higher the beginning of the high rise contour, the more surprised it sounded to subjects. The same logic was applied in the analysis of hyperarticulated vowel formant targets here in Russian unstressed syllables: if two underlyingly distinct vowels that may or may not have been neutralized on the surface move toward the same target under hyperarticulation, then they do not contrast phonologically on the surface. If they move toward different targets, then, modulo the possibility of deep enhancement in some circumstances as yet to be determined, they remain phonologically distinct. It was furthermore argued that this logic may be fruitfully applied in cases of so-called incomplete neutralization. 5 Conclusion I have argued in the above that data from the hyperarticulation of unstressed vowels in Russian provides support for the conclusions of Barnes [2006] regarding the nature of the vowel reduction system in that language. This analysis holds that Russian does indeed have two vowel reduction patterns, as traditional analyses contend. The difference between them, however, is not, as previous analyses have held, a matter of “degree”, such that a more moderate process occurs in one set of unstressed environments while a more extreme process occurs in another, complementary set. Rather, both processes apply equally in all unstressed syllables. One of these is a categorical, phonological neutralization process preventing underlying mid vowels from surfacing as such in unstressed syllables. The second pattern is gradient and phonetic, a continuous function relating vowel height to vowel duration. Where vowel durations are typically shorter (e.g., the second pretonic syllable), duration-dependent undershoot causes significant raising of phonological low vowels toward schwa. Where vowel durations are longer (e.g., in stressed syllables, first pretonic syllables, absolute word-initial or phrase-final position, in hiatus, at slower speech rates, or under hyperarticulation in second pretonic syllables), undershoot decreases, allowing closer approximation of low vowel targets in proportion to the degree of additional duration involved. The impressionistic consequence of this gradient mitigation of undershoot in certain contexts is what has previously traditionally described as failure to reduce to schwa in specific exceptional positions. The results of this study are important both for the analysis of Russian, and for theories of vowel reduction more generally. In particular, the combined results of this study and the speech rate study of Barnes [2006] raise two serious problems for theories of vowel reduction that take duration-dependent undershoot, with its concomitant compression of the vowel space, to be a necessary condition for synchronic phonological vowel reduction to occur. The first problem, the well-known stabilization problem, is tractable. The second problem, I believe, is much less so; there are documented instances, such as the first pretonic syllable in Moscow Russian, in which phonological neutralizations take place in unstressed syllables even in the systematic absence of any significant undershoot or vowel space compression. This study also identifies several promising areas for future research, including the role of increased articulatory effort as a co-predictor along with duration of F1 in unstressed syllables of Russian, possibly as a matter of interspeaker variation, or contextual variation, or both. The need for a closer study of subtle regional differences in the realization of unstressed vowels in Russian was also underscored. Finally, hyperarticulation was shown to be a valuable tool for the investigation of phonological category membership, and possibly for distinguishing the activities of categorical, phonological processes from gradient phonetic ones more generally. The notions of deep and shallow enhancement of contrast were introduced and shown to have serious implications for the analysis of certain perennially recalcitrant questions in the phonetics-phonology interface, incomplete neutralization in particular. Acknowledgments Many thanks to Edward Flemming, Andrew Garrett, Larry Hyman, Sharon Inkelas, Dasha Kavitskaya, John Kingston, Jaye Padgett, Donca Steriade, and Alan Timberlake for helpful comments at various stages of this project. Thanks too to audiences at the 2004 Annual Meeting of the LSA, and the University of Massachusetts, and to numerous helpful Russian speakers in the Boston area for their advice and/or participation in this study. Appendix Word list (all in citation form unless otherwise noted) SECOND PRETONIC UNDERLYING /a/ Orthography IPA Gloss багажа ba ‘baggage’ (gen.sg.) батальон btlon ‘battalion’ кабаре kbre ‘cabaret’ лаборант lbrant ‘lab worker’ паковать pkvat ‘to pack’ патовать ptvat ‘to stalemate’ табачок tbtok ‘tobacco’ (dim.) таковой tkvoj ‘such’ (adj.) лапоток lptok ‘bast sandal’ (dim.) SECOND PRETONIC UNDERLYING /o/ Orthography IPA Gloss богатырь btr ‘bogatyr’ боковой bkvoj ‘side’ (adj.) годовой dvoj ‘year’ (adj.) доказать dkzat ‘to prove’ подошёл pdol ‘approached’ (m.sg.) погадал pdal ‘told fortunes’ (m.sg.) потакать ptkat ‘to indulge’ локоток lktok ‘elbow’ (dim.) лопотать lptat ‘to blather, mutter’ References Alderete J.: Faithfulness to prosodic heads. (unpublished manuscript, University of Massachusetts, Amherst. Rutgers Optimality Archive #94-0000, 1995). Avanesov R.I.: Russkoe literaturnoe proiznoshenie. (Moskva: Prosveshchenie 1972). Barnes J.: Positional Neutralization: a phonologization approach to typological patterns; unpublished Ph.D. diss. University of California, Berkeley (2002). Barnes J.: Strength and weakness at the interface: positional neutralization in phonetics and phonology. (Berlin/New York: Mouton de Gruyter 2006). Beckman J.N.: Positional faithfulness; unpublished PhD diss. University of Massachusetts, Amherst (1998). Beckman M.E.; Edwards J.: The phonological domains of final lengthening. Ohio State University Working Papers in Linguistics 35:167-176 (1987). Beckman M.E.; Edwards J.; Fletcher J.: The articulatory kinematics of final lengthening. Journal of the Acoustical Society of America 89:369-382 (1991). Boersma P.; Weenink D.: Praat: doing phonetics by computer (Version 4.0.2)[Computer program]. (2004). http://www.praat.org/ Bondarko L.V.: Zvukovoi stroi sovremennogo russkogo iazyka. (Moskva: Prosveshchenie 1977). Bradlow A.R.; Kraus N.; Hayes E.: Speaking clearly for learning-impaired children: Sentence perception in noise. Journal of Speech, Language, and Hearing Research 46:80-97 (2003). Bybee J.: Phonology and language use. (Cambridge: Cambridge University Press 2001). Byrd D.: Articulatory vowel lengthening and coordination at phrasal junctures. Phonetica 57:316 (2000). Byrd D.; Krivokapić J.; Lee S.: How far, how long: On the temporal scope of prosodic boundary effects. Journal of the Acoustical Society of America 120:1589-1599 (2006). Byrd D.; Saltzman E.: The elastic phrase: Modeling the dynamics of boundary-adjacent lengthening. Journal of Phonetics 31:149-180 (2003). Cambier-Langeveld T.: The domain of final lengthening in the production of Dutch. Linguistics in the Netherlands 14:13-24 (1997). Cambier-Langeveld T.: The interaction between final lengthening and accentual lengthening. Linguistics in the Netherlands 16:13-25 (1999). Casali R.: Labial opacity and roundness harmony in Nawuri. Natural Language and Linguistic Theory 13:649-663 (1995). Cho T.: The effects of prosody on articulation in English. (New York: Routledge 2001/2002). Cho T.: Prosodic strengthening and featural enhancement: Evidence from acoustic and articulatory realizations of /a,i/ in English. Journal of the Acoustical Society of America 117:3867-3878 (2005). Cho T.; Jun S.-A.: Domain-initial strengthening as enhancement of laryngeal features: Aerodynamic evidence from Korean. UCLA Working Papers in Phonetics 99:57-79 (2000). Cho T.; Keating P.: Articulatory and acoustic studies of domain-initial strengthening in Korean. UCLA Working Papers in Phonetics 97:100-139 (1999). Cho T.; McQueen J.M.: Prosodic influences on consonant production in Dutch: Effects of prosodic boundaries, phrasal accent and lexical stress. Journal of Phonetics 33:121-157 (2005). Cho T.; McQueen J.M.; Cox E.A.: Prosodically driven phonetic detail in speech processing: The case of domain-initial strengthening in English. Journal of Phonetics (in press). Cohn A.: Phonetic and phonological rules of nasalization; unpublished Ph.D. diss. University of California, Los Angeles (1990). Crosswhite K.: Vowel reduction in Optimality Theory. (New York: Routledge 2001). Crosswhite K.: Vowel reduction. In Hayes, B.; Kirchner, R.; Steriade, D., (Eds.) Phonetically Based Phonology, pp 191-231. (Cambridge: Cambridge University Press 2003). Curl T.S.: The phonetics of sequence organization: an investigation of lexical repetition in otherinitiated repair sequences in American English; unpublished Ph.D. diss. University of Colorado (2002). Czekman W.; Smułkowa E.: Fonetyka i fonologia języka Białoruskiego z elementami fonetyki i fonologii ogólnej. (Warszawa: Państwowe Wydawnictvo Naukowe 1988). de Jong K.: The supraglottal articulation of prominence in English: Linguistic stress as hyperarticulation. Journal of the Acoustical Society of America 97:491-504 (1995). de Jong K.: Stress, lexical focus, and segmental focus in English: Patterns of variation in vowel duration. Journal of Phonetics 32:493-516 (2004). Dilley L.; Shattuck-Hufnagel S.; Ostendorf M.: Glottalization of word-initial vowels as a function of prosodic structure. Journal of Phonetics 24:423-444 (1996). Dinnsen D.A.; Charles-Luce J.: Phonological neutralization, phonetic implementation, and individual differences. Journal of Phonetics 12:49-60 (1984). Dinnsen D.A.; Garcia-Zamor M.: Three degrees of vowel length in German. Journal of Linguistics 4:111-126 (1971). Engstrand O.: Articulatory correlates of stress and speaking rate in Swedish VCV utterances. Journal of the Acoustical Society of America 83:1863-1875 (1988). Ernestus M.; Baayen H.: The functionality of incomplete neutralization in Dutch: The case of past-tense formation. In Goldstein, M.L.; Whalen, D.H.; Best, C., (Eds.) Laboratory Phonology 8: Varieties of phonological competence, pp 27-49. (Berlin/New York: Mouton de Gruyter 2006). Ferguson S.H.: Vowels in clear and conversational speech: Talker differences in acoustic features and intelligibility for normal-hearing listeners; unpublished Ph.D. diss. Indiana University (2003). Fidelholtz J.L.: Word frequency and vowel reduction in English. CLS 11:200-213 (1975). Flemming E.: Auditory representations in phonology. (New York: Routledge 1995/2002). Flemming E.: Scalar and categorical phenomena in a unified model of phonetics and phonology. Phonology 18:7-44 (2001). Flemming E.: A phonetically-based model of phonological vowel reduction. (unpublished manuscript, MIT. 2005). Fodor J.A.: The modularity of mind. (Cambridge, MA: MIT Press 1983). Fougeron C.; Keating P.: Articulatory strengthening at edges of prosodic domains. Journal of the Acoustical Society of America 101:3728-3740 (1997). Fourakis M.; Iverson G.: On the 'incomplete neutralization' of German final obstruents. Phonetica 41:140-149 (1984). Gafos A.I.: Dynamics in grammar: Comment on Ladd and Ernestus & Baayen. In Goldstein, M.L.; Whalen, D.H.; Best, C., (Eds.) Laboratory Phonology 8: Varieties of phonological competence, pp 51-79. (Berlin/New York: Mouton de Gruyter 2006). Gussenhoven C.; Rietveld T.: The behavior of H* and L* under variations in pitch range in Dutch rising contours. Language and Speech 43:183-203 (2000). Halle M.: The sound pattern of Russian. (The Hague: Mouton 1959). Halle M.; Vergnaud J.R.: An essay on stress. (Cambridge: MIT Press 1987). Hayes B.: Phonetically-driven phonology: The role of Optimality Theory and inductive grounding. In Darnell, M.; Moravcsik, E.; Noonan, M.; Newmeyer, F.; Wheatly, K., (Eds.) Functionalism and formalism in linguistics, volume 1: General papers, pp 243-285. (Amsterdam: John Benjamins 1999). Hayes B.; Steriade D.: Introduction: The phonetic bases of phonological markedness. In Hayes, B.; Kirchner, R.; Steriade, D., (Eds.) Phonetically based phonology, pp 1-33. (Cambridge: Cambridge University Press 2003). Herrick D.: An acoustic analysis of phonological vowel reduction in six varieties of Catalan; unpublished Ph.D. diss. University of California, Santa Cruz (2003). Hooper J.B.: Word frequency in lexical diffusion and the source of morphophonological change. In Christie, W., (Ed.) Current progress in historical linguistics, pp 95-105. (Amsterdam: North Holland 1976). Hubbard K.: Duration in moraic theory; unpublished Ph.D. diss. University of California, Berkeley (1994). Jassem W.: The phonology of Polish stress. Word 15:252-269 (1959). Johnson K.; Flemming E.; Wright R.: The hyperspace effect: Phonetic targets are hyperarticulated. Language 69:505-528 (1993). Johnson K.; Martin J.: Acoustic vowel reduction in Creek: Effects of distinctive length and position in the word. Phonetica 58:81-102 (2001). Jones D.; Ward D.: The phonetics of Russian. (Cambridge: Cambridge University Press 1969). Jurafsky D.; Bell A.; Gregory M.; Raymond W.D.: Probabilistic relations between words: Evidence from reduction in lexical production. In Bybee, J.; Hopper, P., (Eds.) Frequency and the emergence of linguistic structure, pp 229-254. (Amsterdam: John Benjamins 2001). Keating P.: The phonology-phonetics interface. In Newmeyer, F., (Ed.) Linguistics: The Cambridge survey, Volume I: Grammatical theory, pp 281-302. (Cambridge: Cambridge University Press 1988). Keating P.: The phonology-phonetics interface. In Kleinhenz, U., (Ed.) Interfaces in phonology, pp 262-278. (Berlin: Akademie Verlag 1996). Keating P.; Cho T.; Fougeron C.; Hsu C.-S.: Domain-initial strengthening in four languages. UCLA Working Papers in Phonetics 97:139-151 (1999). Keating P.; Wright R.; Zhang J.: Word-level asymmetries in consonant articulation. UCLA Working Papers in Phonetics 97:157-173 (1999). Kirchner R.: An effort-based approach to consonant lenition. (New York: Routledge 1998/2001). Klatt D.: Vowel lengthening is syntactically determined in a connected discourse. Journal of Phonetics 3:129-140 (1975). Kuehn D.P.; Moll K.L.: A cineradiographic study of VC and CV articulatory velocities. Journal of Phonetics 4:303-320 (1976). Labov W.: Principles of Linguistic Change. Volume 1: Internal factors. (Oxford: Blackwell 1994). Labov W.; Karan M.; Miller C.: Near-mergers and the suspension of phonemic contrast. Language Variation and Change 3:33-74 (1991). Lehiste I.: Suprasegmentals. (Cambridge: MIT Press 1970). Lehiste I.; Ivić P.: Word and sentence prosody in Serbocroatian. (Cambridge: MIT Press 1986). Liberman M.Y.; Pierrehumbert J.B.: Intonational invariance under changes in pitch range and length. In Aronoff, M.; Oehrle, R., (Eds.) Language sound structure, pp 157-233. (Cambridge, MA: MIT Press 1984). Lightner T.: Segmental phonology of Modern Standard Russian; unpublished Ph.D. diss. MIT (1965). Lightner T.: Problems in the theory of phonology, Vol. 1: Russian phonology and Turkish phonology. (Edmonton: Linguistic Research, Inc. 1972). Liljencrants J.; Lindblom B.: Numerical simulation of vowel quality systems: the role of perceptual contrast. Language 48:839-862 (1972). Lindblom B.: A spectrographic study of vowel reduction. Journal of the Acoustical Society of America 37:1773-1781 (1963). Lindblom B.: Explaining phonetic variation: A sketch of the H&H theory. In Hardcastle, W.J.; Marchal, A., (Eds.) Speech production and speech modelling, pp 403-439. (Dordrecht: Kluwer Academic Publishers 1990). Liphola M.M.: Aspects of the phonology and morphology of Shimakonde; unpublished Ph.D. diss. Ohio State University (2001). Liu S.; Del Rio E.; Bradlow A.R.; Zeng F.G.: Clear speech perception in acoustic and electrical hearing. Journal of the Acoustical Society of America 116:2374-2383 (2004). Matusevič M.I.: Sovremennyi russkii iazyk: Fonetika. (Moskva: Prosveshchenie 1976). Max L.; Onghena P.: Some issues in the statistical analysis of completely randomized and repeated measures designs for speech, language, and hearing research. Journal of Speech, Language, and Hearing Research 42:261-270 (1999). Moon S.-J.: An acoustic and perceptual study of formant undershoot in clear and citation-form speech; unpublished Ph.D. diss. University of Texas, Austin (1991). Moon S.-J.; Lindblom B.: Interaction between duration, context, and speaking style in English stressed vowels. Journal of the Acoustical Society of America 96:40-55 (1994). Myers S.: Boundary disputes: The distinction between phonetic and phonological sound patterns. In Burton-Roberts, N.; Carr, P.; Docherty, G., (Eds.) Phonological knowledge: Conceptual and empirical issues, pp 245-272. (Oxford: Oxford University Press 2000). Nord L.: Vowel reduction–Centralization or contextual assimilation? In Fant, G., (Ed.) Proceedings of the speech communication seminar, Stockholm. Aug. 1-3, 1974, pp 149-154. (Stockholm: Almqvist & Wiksell 1975). Nord L.: Acoustic studies of vowel reduction in Swedish. Proceedings of the eleventh international congress of phonetic sciences, August 1-7, 1987, Tallinn, Estonia, USSR, pp 157160. (Tallinn: Academy of Sciences of the Estonian S.S.R. 1987). Nowak P.M.: Vowel reduction in Polish; unpublished Ph.D. diss. University of California, Berkeley (2006). Ohala J.J.: The listener as a source of sound change. In Masek, C.S.; Hendrick, R.A.; Miller, M.F., (Eds.) Papers from the parasession on language and behavior, Chicago Linguistic Society, pp 178-203. (Chicago: CLS 1981). Oller D.K.: The duration of speech segments: The effect of position in utterance and word length. Journal of the Acoustical Society of America 54:1235-1273 (1973). Padgett J.: Contrast dispersion and Russian palatalization. In Hume, E.; Johnson, K., (Eds.) The role of speech perception in phonology, pp 187-218. (San Diego, CA: Academic Press 2001). Padgett J.; Tabain M.: Adaptive dispersion theory and phonological vowel reduction in Russian. Phonetica 62:14-54 (2005). Picheny M.A.; Durlach N.I.; Braida L.D.: Speaking clearly for the hard of hearing. II. Acoustic characteristics of clear and conversational speech. Journal of Speech and Hearing Research 29:434-446 (1986). Pierrehumbert J.B.: Word-specific phonetics. In Gussenhoven, C.; Warner, N., (Eds.) Laboratory Phonology VII, pp 101-139. (Berlin/New York: Mouton de Gruyter 2002). Port R.F.; Crawford P.: Incomplete neutralization and pragmatics in German. Journal of Phonetics 17:257-282 (1989). Port R.F.; O'Dell M.L.: Neutralization of syllable-final voicing in German. Journal of Phonetics 13:455-471 (1985). Prince A.; Smolensky P.: Optimality Theory: Constraint ranking in Generative Grammar. (Oxford: Blackwell 1993/2004). Savitska I.; Spasov L.: Fonologija na sovremeniot makedonski standarden jazik: segmentalna i suprasegmentalna. (Skopje: Detska Radost 1997). Shcherba L.V.: Russkie glasnye v kachestvennom i kolichestvennom otnoshenii. (Sankt Peterburg 1912). Slowiaczek L.M.; Dinnsen D.A.: On the neutralizing status of Polish word-final devoicing. Journal of Phonetics 13:325-341 (1985). Smiljanić R.; Bradlow A.R.: Production and perception of clear speech in English and Croatian. Journal of the Acoustical Society of America 118:1677-1688 (2005). Smith J.: Phonological augmentation in prominent positions; unpublished Ph.D. diss. University of Massachusetts, Amherst (2002). Solé M.-J.: Phonetic and phonological processes: the case of nasalization. Language and Speech 35:29-43 (1992). Steriade D.: Positional neutralization and the expression of contrast. (unpublished manuscript, University of California, Los Angeles. 1995). Steriade D.: Phonetics in phonology: The case of laryngeal neutralization. In Gordon, M., (Ed.) Papers in Phonology 3. (UCLA Working Papers in Linguistics 2 1999). Steriade D.: Paradigm uniformity and the phonetics-phonology boundary. In Broe, M.B.; Pierrehumbert, J.B., (Eds.) Papers in Laboratory Phonology V: Acquisition and the lexicon, pp 313-334. (Cambridge: Cambridge University Press 2000). Stojkov S.: Bâlgarska dialektologija. (Sofia: Izdatelstvo na BAN 1962/1993). Tiffany W.R.: Nonrandom sources of variation in vowel quality. Journal of Speech and Hearing Research 2:303-317 (1959). Tilkov D.: Gramatika na Sâvremennija Bâlgarski Ezik: Tom 1/ Fonetika. (Sofia: Izdatelstvo na BAN 1982). Timberlake A.: A reference grammar of Russian. (Cambridge: Cambridge University Press 2004). Trubetzkoy N.S.: Principles of phonology. (Berkeley: University of California Press 1969). Uchanski R.M.; Choi S.S.; Braida L.D.; Reed C.M.; Durlach N.I.: Speaking clearly for the hard of hearing. IV. Further studies of the role of speaking rate. Journal of Speech and Hearing Research 39:494-509 (1996). van Son R.J.J.H.; Pols L.C.W.: Formant frequencies of Dutch vowels in a text, read at normal and fast rate. Journal of the Acoustical Society of America 88:1683-1693 (1990). van Son R.J.J.H.; Pols L.C.W.: Formant movements of Dutch vowels in a text, read at normal and fast rate. Journal of the Acoustical Society of America 92:121-127 (1992). Ward D.: Unaccented vowels in Russian. Russian Linguistics 2:91-104 (1975). Warner N.; Jongman A.; Sereno J.; Kemps R.: Incomplete neutralization and other sub-phonemic durational differences in production and perception: Evidence from Dutch. Journal of Phonetics 32:251-276 (2004). Wightman C.; Shattuck-Hufnagel S.; Ostendorf M.; Price P.J.: Segmental durations in the vicinity of prosodic phrase boundaries. Journal of the Acoustical Society of America 91:17071717 (1992). Yu A.C.L.: In defense of phonetic analogy: evidence from tonal morphology. Paper presented at the 2006 Annual Meeting of the Linguistics Society of America, Albuquerque, January 5-8. (2006). Zhang J.: The effects of duration and sonority on contour tone distribution–typological survey and formal analysis; unpublished Ph.D. diss. University of California, Los Angeles (2001). Zipf G.K.: Relative frequency as a determinant of phonetic change. Harvard Studies in Classical Philology 15:1-95 (1929). Zlatoustova L.V.: Foneticheskaia struktura slova v potoke rechi. (Kazan': Izdatel'stvo Kazanskogo Universiteta 1962). Zoll C.: Positional asymmetries and licensing. (unpublished manuscript, MIT. Rutgers Optimality Archive #282-0998, 1998). Zsiga E.C.: Features, gestures and the temporal aspects of phonological organization; unpublished Ph.D. diss. Yale University (1993). Table 1. Mean first formant values in Hz and standard deviations Speaker RR TM DT M All 2nd pretonic /a/ /o/ 482 490 (54) (49) 443 449 (36) (42) 492 506 (42) (63) 484 483 (59) (50) 475 481 (52) (55) 1st pretonic /a/ 734 (53) 713 (79) 641 (34) 734 (52) 705 (68.5) stressed /o/ 720 (45) 695 (90) 640 (32) 742 (70) 694 (73) /a/ 723 (53) 691 (88) 637 (32) 784 (42) 700 (78) /o/ 450 (25) 444 (38) 455 (30) 503 (34) 461 (39) Table 2. Mean vowel durations in ms. and standard deviations Speaker RR TM DT M All 2nd pretonic /a/ /o/ 21.7 26.2 (7.2) (7.6) 24.2 26.6 (7.5) (9.4) 26.3 32.6 (10) (10.6) 28.1 31.6 (10.4) (9.9) 25.2 28.9 (9.2) (9.7) 1st pretonic /a/ 65.9 (9.9) 81.1 (19.3) 58.2 (13.6) 70.9 (14.5) 69 (16.9) /o/ 69.1 (11.6) 80.4 (18.5) 60.7 (13.5) 70.9 (13.7) 69.9 (16.2) stressed /a/ 85.9 (14) 91.7 (25.7) 72.6 (13.2) 75.2 (11.4) 81.8 (19.3) /o/ 77.6 (18.1) 81.4 (32) 69.3 (12.6) 68.7 (15.4) 74.5 (21.5) Table 3. Mean first formant values in Hz and standard deviations Speaker RR TM DT M All 2nd pretonic /a/ 548 (57) 581 (153) 618 (110) 553 (81) 576 (110) /o/ 529 558 574 547 552 (55) (159) (110) (92) (109) Table 4. Mean vowel durations in ms and standard deviations Speaker RR TM DT M All 2nd pretonic /a/ 32.5 (7.7) 71.3 (54.5) 38.2 (10.3) 35.5 (16.1) 44.9 (33.4) /o/ 33.4 75.5 39.5 40.9 47.1 (8.6) (55.8) (11) (22.7) (34.5) Table 5. ANOVA results assessing the effect of underlying vowel on F1 for each speaker SPEAKER RR TM DT M FIRST FORMANT F(1, 49) = 1.565, p = .217 F(1, 51) = .291, p = .592 F(1, 50) = 2.063, p = .157 F(1, 51) = .065, p = .799 DURATION F(1, 49) = .153, p = .697 F(1, 51) = .076, p = .784 F(1, 50) = .185, p = .669 F(1, 51) = 1.005, p = .321 Table 6. Mean and Standard Deviation for F1 and duration of /o/: prefix vs. root /o/ Mean F1 Duration Root 560 (122) 49.8 (35) Prefix 542 (92) 43.9 (33.9) Table 7. Linear regression models for F1 as a function of duration for all speakers in the hyperarticulation study. SPEAKER RR TM DT M LINEAR REGRESSION MODEL y = 4.0109x + 405.48 y = 2.6285x + 377.08 y = 7.4933x + 304.66 y = 3.6662x + 410.1 R-SQUARED R2 (51) = 0.3391; p < .001 R2 (53) = 0.8608; p < .001 R2 (52) = 0.5067; p < .001 R2 (53) = 0.7077; p < .001 Figure 1. Russian second pretonic /a/, /o/, all speakers: F1 vs. Duration (after Barnes [2006: 58]) 800 /o/: R2 = 0.2468 First formant in Hz 700 /a/: R2 = 0.3584 600 /a/ 500 /o/ Linear (/a/) 400 Linear (/o/) 300 200 100 0 0 10 20 30 40 Vowel durations in ms. 50 60 70 Figure 2. Unstressed /a, o/ for all speakers, contexts combined Speaker RR Speaker TM 1000 1000 900 900 800 800 2 pretonic /a, o/ 500 400 700 F1 in Hz 1 pretonic /a, o/ 600 2 pretonic /a, o/ 500 400 300 200 200 100 100 0 1 pretonic /a, o/ 600 300 0 0 20 40 60 80 100 0 20 40 60 Duration in ms 80 100 120 140 Duration in ms Speaker DT Speaker M 800 1000 700 900 800 600 1 Pretonic /a, o/ 500 2 pretonic /a, o/ 400 300 700 F1 in Hz F1 in Hz F1 in Hz 700 1 pretonic /a, o/ 600 2 pretonic /a, o/ 500 400 300 200 200 100 100 0 0 0 20 40 60 Duration in ms 80 100 0 20 40 60 Duration in ms 80 100 120 Figure 3. Mean first and second formant values, pooled across speakers, for stressed /a/ and /o/, and three repetitions, one casual and two hyperarticulated, of second pretonic underlying /a/ and /o/. Figure 4. Vowel height as a function of phonetic duration, speaker by speaker Speaker RR: Second Pretonic /a/ and /o/ Speaker TM: Second Pretonic /a/ and /o/ 700 1200 600 1000 F1 in Hz F1 in Hz 500 400 300 200 800 600 400 200 100 0 0 0 10 20 30 40 50 60 0 50 100 Duration in ms F1 in Hz F1 in Hz 1000 900 800 700 600 500 400 300 200 100 0 20 40 Duration in ms 200 250 Speaker M: Second Pretonic /a/ and /o/ Speaker DT: Second Pretonic /a/ and /o/ 0 150 Duration in ms 60 80 1000 900 800 700 600 500 400 300 200 100 0 0 50 100 Duration in ms 150 Figure 5. Speaker TM: Underlying /a/ and /o/ coded for repetition number. 1000 900 800 F1 in Hz 700 600 Repetition 1 500 Repetition 2 400 Repetition 3 300 200 100 0 0 50 100 150 Duration in ms 200 250 Figure 6. Speaker RR: Underlying /a/ and /o/ coded for repetition number. 700 600 F1 in Hz 500 400 Repetition 1 300 Repetition 3 Repetition 2 200 100 0 0 20 40 Duration in ms 60 Figure 7. Vowel height as a function of duration. All speakers: Second Pretonic /a/ and /o/. 1200 F1 in Hz 1000 800 600 400 200 0 0 50 100 150 Duration in ms 200 250 Figure 8. All speakers: Second Pretonic /a/ and /o/, Standardized Scores. 5.00 4.00 F1 in SD units 3.00 2.00 1.00 -2.00 0.00 -1.00 0.00 -1.00 1.00 2.00 3.00 -2.00 Duration in SD units 4.00 5.00 6.00 Figure 9. Second Pretonic /a/ and /o/, Speaker DT excluded. 1000 900 800 F1 in Hz 700 600 500 400 300 200 100 0 0 50 100 150 Duration in ms 200 250 Figure 10. All speakers: second pretonic /a/ and /o/. Influential cases removed. 1000 900 800 F1 in Hz 700 600 500 400 300 200 100 0 0 20 40 60 80 Duration in ms 100 120 140 Figure 11. Comparison of the spread of duration and F1 values in the hyperarticulation experiment and the speech rate study of Barnes [2006]. Boxes represent the interquartile ranges of values for each speaker in the two studies, with the median indicated by a heavy horizontal line across each box. Whiskers denote the full range of values obtained, excluding outliers. 1 50 9 00 8 00 1 25 F1 in Hz Duration in ms 7 00 1 00 75 6 00 5 00 50 4 00 Experiment 25 Experiment 3 00 Spee ch Ra te 0 Hy pera rticul ation DT M RR SPEAKER TM Spee ch Ra te 2 00 Hy pera rticul ation DT M RR SPEAKER TM