Search

Subjects judged the similarities among a set of American English vowels (/i,I,q,1,v,a,c,o,0,u/) presented in isolation or in a /dVd/ consonantal frame. Individual differences scaling was employed to analyze these similarities data for each of the conditions separately and for the two conditions combined. In all cases, perceptual dimensions corresponding to the advancement, height, and tenseness vowel features were recovered. Given the determinacy of individual differences scaling, this finding is taken to provide strong evidence for the perceptual significance of those features. The perceptual dimensions are considered in relation to various acoustic parameters of the stimuli employed in this study. They are also considered in relation to perceptual dimensions that have been observed in other vowel scaling studies.

It is commonly assumed that a listener identifies an unknown talker's vowels in terms of the relation between their acoustic structure and that of other vowels produced by the same talker: the listener “normalizes” the variant acoustic signals on the basis of information about their source (vocal‐tract characteristics). In order to measure the vowel ambiguity attributable to talker differences, we compared identification of nine vowels in /p—vowel—p/ environment produced by a single talker with the same vowels produced by 15 talkers (men, women, children) who differed widely in vocal‐tract sizes and characteristic fundamental frequency. Mean error rates were 9% and 17%, respectively. Acoustic differences attributable to talker accounted for fewer than half of the perceptual errors. Lieberman [Am. Anthropologist (1972)] has proposed that the point vowels, /i, a, u/, are the “primary calibrating signals” by which the listener determines the size of the talker's vocal tract and normalizes the acoustic signal. In a second study, one group of listeners received each test word preceded by syllables containing /i, a, u/ and another group heard the test words preceded by nonpoint vowels /I, ae, Λ/. There was no significant overall reduction in errors of vowel identification for words preceded either by point vowels or by nonpoint vowels. Although perception of certain vowels improved when preceding syllables were present, perception of other vowels deteriorated.

Listeners make remarkably few errors in identification of randomly presented tokens of vowels produced by a panel of talkers in fixed consonantal environments. The present study was designed to investigate the sources of perceptual confusions among vowels and to determine whether or not the effects of talker differences on vowel identification are independent of consonantal environment. Tokens of nine American English vowels were spoken in three contexts: (1) in isolation, (2) in a variable consonantal environment (/C1—vowel—C2/, where C1,2 = /b, p, d, t, g, k/), and (3) in a fixed consonantal environment (/p—vowel—p/). Test lists were produced by three representative single talkers (man, woman, child) and by a panel of talkers who differed in vocal‐tract size and characteristic fundamental frequency. Whether produced by the panel or by a single talker, isolated vowels were significantly more often misidentified than were vowels in consonantal environments. Reliable differences in identification of medial...

The identifiability of isolated vowels (/V/) was compared to that of vowels in consonantal context (/pVp/) when subjects performed a monitoring task. On successive blocks of trials in a test series, the subjects listened for instances of one or another of nine monophthongal vowels (/i,I,q,1,v,a,c,U,u/) and identified each test item as being an instance or not. On average, resulting false alarm errors occurred significantly less often in the /pVp/ condition, consistent with the previous finding that vowel perception may be aided by consonantal context. This beneficial effect of context was found to be restricted to the class of open vowels, however, with perception of the close vowels being somewhat hindered by context. The error data for misses also showed an interaction between context and vowel height. Various accounts of the interaction are considered.

Variations in syllable stress create a constancy problem for vowel perception owing to the accompanying variations in formant structure. An earlier study [R. R. Verbrugge and D. P. Shankweiler, J. Acoust. Soc. Am. 61, 539 (1977)], failed to demonstrate perceptual “compensation” for stress changes, possibly because variation in the prosody of a carrier phrase was not effective in altering the perceived stress of the target syllables. In the present experiment, stress was varied by altering the acoustic structure of the target syllable itself. A male talker produced a set of carrier sentences containing either stressed or destressed /pVp/ syllables (V = /i, , e, ae, a, , , ν, u/). Two types of synthetic analogs were created using an OVE‐III synthesizer: (a) original formant analogs, matched in pitch, amplitude, duration, and formant frequencies to the original syllables, and (b) interchanged‐formant analogs, preserving the original prosody, but containing formants taken from the syllable of opposite stress. The synthetic syllables were embedded in the original carrier sentences for presentation to listeners. Vowel identification was influenced by perceived stress: errors on stressed formants increased somewhat when the prosody was inappropriately destressed, while errors on destressed formants were paradoxically lower in the presence of stressed prosody. The results demonstrate that vowel identity is perceived in relation to stress. Comparison of error patterns for stress‐interchanged and rate‐interchanged syllables suggests that the two transformations are not equivalent. [Research supported by NICHD.]

A male and a female talker each produced a set of /bVb/ syllables (V = /i,I,e,e,ae,a,ʌ,ɔ,o,ʋ,u/) at metronome‐controlled rates. The central 60% of the voiced regions of these syllables was edited out, producing single‐talker silent‐center syllables [cf. W. Strange, J. J. Jenkins, and T. R. Edman, J. Acoust. Soc. Am. 61, S39 (1977)]. Comparable “hybrid” silent‐center syllables were created by appending the final 20% of one talker's syllable to the initial 20% of the other talker's production of the same syllable, again with a 60% silent region intervening. Vowel identification errors were relatively low for both the single‐talker (22.4%) and the hybrid (27.0%) silent‐center syllables, though errors in each case exceeded those found for the original (unedited) syllables, though errors in each case exceeded those found for the original (unedited) syllables (9.3%). Hybrids were generally heard as coherent syllables from a single source; the error pattern for the hybrids could not be interpreted as a joint function of errors on the initial pieces (55.4% on average) and the final pieces (72.9%). Results indicate that considerable information for vowels is contained in the dynamic regions early and late in a syllable. Moreover, they suggest that this information does not simply subserve computation of an underlying normalized target. The information may specify properties of production dynamics that are invarient for a vowel for all talkers of a dialect. To the extent that such information is available, normalization in perception is unnecessary. [Research supported by NICHD.]

Quasi‐steady‐state vowels which attain representative formant frequency values for adult male speakers of General American might be regarded as good approximations to canonical form. Results of experiments on vowel perception do not support this assumption. Two speakers produced tokens of nine different vowels in isolation. Separate listening tests were made by random orders of six tokens of each vowel. A second pair of tests was created by digitizing the speech signals, abstracting a single pitch pulse (at the zero crossings) from each vowel center, and iterating to produce sets of pseudovowels matched in duration to the parent syllables. Listeners' judgments contained nearly twice as many errors for iterated pseudovowels than for the unedited natural versions. A third experiment with OVE‐synthesized vowels yielded error rates comparable to those for iterated pseudovowels, not to their natural counterparts. We conclude that natural vowels contain sources of information not captured by target formant freq...

Previous studies comparing the perception of isolated vowels with that of vowels produced in some consonantal context have generally focused on similarities and differences in the average performance of subjects in the two conditions. In this investigation, we wanted to look carefully at individual differences in vowel perception both within and between these conditions. We did so with the aid of a nonmetric individual differences scaling procedure developed by Takane et al. [Psychometrika 42, 7–67 (1977)]. The variance common to all subjects (in the two conditions combined) was modeled in a single multidimensional space, and the individual differences were represented as weights (or saliences) that each subject attached to the dimensions of the space. The data were similarity judgments collected (with the method of triadic comparisons) for a set of isolated vowels and for a set of vowels in context recorded by the same talker. The perceptual space had three dimensions. The first of these corresponded clo...

Earlier studies have shown that vowels spoken in isolation tend to be poorly perceived, even when they are produced by phonetically trained talkers. Listeners, however, generally make remarkably few errors in identification of vowels in CVC environment, even when each syllable is uttered by a different talker. Sets of nine American English vowels were spoken by a panel of talkers: in isolation and in a /p‐p/ environment. Measurements of the first three formant frequencies were obtained from spectrograms. Listening tests were made up by randomizing talkers and tokens, and these were presented to phonetically naive listeners. Percent recognition of the intended vowel (average over vowels) was 83% for the /p‐p/ condition and 58% for the isolated condition. When the formant frequencies of each talker's isolated and medial vowels are compared, the values are found to be highly similar. Another explanation must be sought for the perceptual difficulty of isolated steady‐state vowels.

We have previously reported that an individual differences scaling comparison of the perception of isolated vowels and the perception of vowels in a consonantal environment (/dVd/) indicated the following: vowels are processed in a more linguistically appropriate way when they occur in consonantal context [B. Rakerd and R. R. Verbrugge, J. Acoust, Soc. Am. Suppl. 1 71, S76 (1982)]. In the present investigation, the method of that earlier experiment was extended to a study of vowel memory. Subjects were asked to make similarity judgments about their memories of (a) isolated vowels, or (b) vowels in /dVd/ consonantal context, and the resulting similarities matrices were submitted to nonmetric individual differences scaling. The scaling analysis indicated (1) that vowel memory, like vowel perception, is to a measurable degree organized around the features of advancement, height, and tenseness; and (2) that the previously reported perceptual influence of consonantal context is largely to be explained in terms...

Interest in the perceptual equivalence of “static” and “dynamic” stimuli spans both speech and nonspeech domains. Previous investigations of speech stimuli have shown that, given matched sets of three‐formant steady‐state vowels and vowels flanked by consonantal transitions, phonetic equivalence is not defined by equal position along the two continua [B. Lindblom and M. Studdert‐Kennedy, J. Acoust. Soc. Am. 42, 830–843 (1967)]. For judgments of relative pitch, on the other hand, ramp tones are judged to be nearly equivalent in pitch to steady‐state tones at their endpoint frequencies [P. T. Brady et al., J. Acoust. Soc. Am. 33, 1357–1362 (1961)]. The present set of experiments demonstrates this difference in categorization of static and dynamic stimuli using a single type of stimulus. Two sets of sine wave stimuli (FLAT/CONTOUR) were constructed by modeling the frequency and relative amplitude characteristics of the aforementioned speech stimuli. When judged in terms of phonetic categories (/ᴜ/ vs /ɪ/), significantly greater proportion of /ᴜ/ judgments were observed for the FLAT than for the CONTOUR continuum. However, when the same two sets of stimuli were judged in terms of relative pitch (HIGH versus LOW), no significant difference in response pattern for the two continua was noted. Thus the experiments demonstrate that two distinct response patterns may be elicited for sine wave stimuli contingent only on the experimenter's instructions. [Work supported by NICHD.]

In rapid speech, acoustic analysis reveals that steady state vowel targets characteristically are not reached. Lindblom and Studdert‐Kennedy [J. Acoust. Soc. Am. 42, 830–843 (1967)] found in an experiment with synthetic speech that listeners showed a shift in the boundary between medial vowels /e/ and /U/ with variations in the rate and direction of formant transitions. Apparently, perceivers compensate for simulated articulatory undershoot by perceptual overshoot. An experiment with natural speech demonstrated shifts in the acoustic criteria listeners employed in vowel recognition as a function of perceived rate of utterance. Nine American English vowels in /p‐p/ environment were produced by a panel of 15 talkers in a fixed sentence frame. The unstressed, rapidly articulated /p‐p/ syllables were excised from the tape recording and assembled into listening tests. In one condition, the /p‐p/ targets were prefaced by three slowly articulated syllables /hi, ha, hu/. In the other test condition, no precursors...

Consonantal environment does not aid vowel identification universally, but preserves and sharpens the information for vowels differentially. In a vowel recognition task, listeners heard one of two test series: /pVp/ syllables or isolated vowels. The test items were spoken by a single male talker and contained a balanced distribution of the vowels /i, ɪ, e, ae, a, ɔ, ʌ, ᴜ, u/. During each block of 90 test trials, listeners targeted for one of the nine vowel types, checking “Yes” when they recognized an instance and “No” otherwise. For both misses and false alarms, there was a strong interaction between vowel type (close versus open) and context type (/pVp/ vs /V/). The open vowel pairs (/ɛ/ae/ , /ʌ/a/) showed consistently superior recognition in the /pVp/ consonantal environment; others did not. Asymmetries in confusion errors were observed for several vowel pairs in both environments. A dynamic articulatory model of the syllable can provide a parsimonious account for these effects. Listeners may be sensitive to acoustic patterning that specifies vowels as dynamic gestures; vowel recognition will vary in precision depending on how the consonantal envirionment structures the dynamic information in sound. [Research supported by NIH grant HD‐01994 to Haskins Laboratories.]