Steady-state Formant Research Articles

Vowels in clear and conversational speech: Within-talker variability in acoustic characteristics

The Ferguson Clear Speech Database was developed for the first author’s doctoral dissertation, which was directed by Diane Kewley-Port at Indiana University. While most studies using the Ferguson Database have examined variability among the 41 talkers, the present investigation considered within-talker differences. Specifically, this study examined the amount of variability each talker showed among the 7 tokens of each of 10 vowels produced in clear versus conversational speech. Steady-state formant frequencies have been measured for 5740 vowels in /bVd/ context using Praat, and a variety of measures of spread will be used to determine variability for each vowel in each speaking style for each talker. Results will be compared to those of the only known previous study that included a sufficiently large number of tokens for this type of analysis, an unpublished thesis from 1980. Based on this study, we predict that within-token variability will be smaller in clear speech than in conversational speech.

Determining the relevance of different aspects of formant contours to intelligibility

Previous studies have shown that “clear” speech, where the speaker intentionally tries to enunciate, has better intelligibility than “conversational” speech, which is produced in regular conversation. However, conversational and clear speech vary along a number of acoustic dimensions and it is unclear what aspects of clear speech lead to better intelligibility. Previously, Kain et al. (2008) showed that a combination of short-term spectra and duration was responsible for the improved intelligibility of one speaker. This study investigates subsets of specific features of short-term spectra including temporal aspects. Similar to Kain’s study, hybrid stimuli were synthesized with a combination of features from clear speech and complementary features from conversational speech to determine which acoustic features cause the improved intelligibility of clear speech. Our results indicate that, although steady-state formant values of tense vowels contributed to the intelligibility of clear speech, neither the steady-state portion nor the formant transition was sufficient to yield comparable intelligibility to that of clear speech. In contrast, when the entire formant contour of conversational speech including the phoneme duration was replaced by that of clear speech, intelligibility was comparable to that of clear speech. It indicated that the combination of formant contour and duration information was relevant to the improved intelligibility of clear speech. The study provides a better understanding of the relevance of different aspects of formant contours to the improved intelligibility of clear speech.

The effect of increased loudness on anticipatory coarticulation of steady-state formants in a vowel-consonant-vowel context for patients with Parkinson’s disease.

The current investigation examines the effect of increased loudness on anticipatory coarticulation of the vowels [i, ɑ, u] on the preceding schwa vowel for patients with Parkinson’s disease. Four speakers, three males and one female with speech severity levels between 8% and 49%, are involved in pre-treatment, treatment, and post-treatment conditions. During treatment, participants use an auditory device which plays speech-babble into one ear while they are speaking. The speech-babble competition results in an automatic loudness increase (Lombard effect) in the talker wearing the device. Participants produced target nonsense CVC words embedded in carrier sentences at comfortable loudness levels and at an average of 3 dB above that level. The steady-state F1, F2, and F3 formant frequencies of the preceding [ə] were analyzed to determine whether increased loudness results in increased anticipatory coarticulation from the [i, ɑ, u] vowels in the CVC words, as treatment progresses. Results using steady-state formant frequency measures from this study will be compared with previous studies that used F2 trajectory measures [Tjaden and Sussman (2006); Weismer et al. (1995); Tjaden (2003)]. (Work supported by NIH 1R01DC009409-01)

Acoustic correlates of clear speech vowel intelligibility for elderly hearing-impaired listeners.

Earlier work on vowels in clear and conversational speech produced by a single talker suggested that the acoustic cues that underlie the superior intelligibility of clear speech for older adults with hearing loss may differ from the cues that young normal-hearing listeners find beneficial. In the present study, clear and conversational vowel stimuli produced by all 41 talkers from the Ferguson Clear Speech Database were presented in a background of 12-talker babble to 40 older adults with mild-to-moderately severe sloping sensorineural hearing loss. Acoustic analyses (vowel duration, steady-state formant frequencies, and measures of dynamic formant movement) were also performed for all stimuli. Vowel intelligibility scores for each talker in each speaking style as well as the magnitude of the clear speech vowel intelligibility effect for each talker will be compared to those obtained from young listeners with normal hearing in a previous study. Regression and other statistical analyses of intelligibility and acoustic data will also be performed to determine the acoustic correlates associated with improved clear speech vowel intelligibility for each listener group. [Work supported by NIHDCD-008886.]

Dynamic spectral structure specifies vowels for children and adults.

When it comes to making decisions regarding vowel quality, adults seem to weight dynamic syllable structure more strongly than static structure, although disagreement exists over the nature of the most relevant kind of dynamic structure: spectral change intrinsic to the vowel or structure arising from movements between consonant and vowel constrictions. Results have been even less clear regarding the signal components children use in making vowel judgments. In this experiment, listeners of four different ages (adults, and 3-, 5-, and 7-year-old children) were asked to label stimuli that sounded either like steady-state vowels or like CVC syllables which sometimes had middle sections masked by coughs. Four vowel contrasts were used, crossed for type (front/back or closed/open) and consonant context (strongly or only slightly constraining of vowel tongue position). All listeners recognized vowel quality with high levels of accuracy in all conditions, but children were disproportionately hampered by strong coarticulatory effects when only steady-state formants were available. Results clarified past studies, showing that dynamic structure is critical to vowel perception for all aged listeners, but particularly for young children, and that it is the dynamic structure arising from vocal-tract movement between consonant and vowel constrictions that is most important.

An acoustic description of the vowels of northern and southern standard Dutch II: Regional varieties

An analysis is presented of regional variation patterns in the vowel system of Standard Dutch as spoken in the Netherlands (Northern Standard Dutch) and Flanders (Southern Standard Dutch). The speech material consisted of read monosyllabic utterances in a neutral consonantal context (i.e., /sVs/). The analyses were based on measurements of the duration and the frequencies of the first two formants of the vowel tokens. Recordings were made for 80 Dutch and 80 Flemish speakers, who were stratified for the social factors gender and region. These 160 speakers were distributed across four regions in the Netherlands and four regions in Flanders. Differences between regional varieties were found for duration, steady-state formant frequencies, and spectral change of formant frequencies. Variation patterns in the spectral characteristics of the long mid vowels /e o ø/ and the diphthongal vowels /ei oey bacwards c u/ were in accordance with a recent theory of pronunciation change in Standard Dutch. Finally, it was found that regional information was present in the steady-state formant frequency measurements of vowels produced by professional language users.

Formant detail needed for vowel identification

Changes in formant frequency over time are important for vowel identification: listeners identify stimuli containing time-varying formants better than stimuli with steady-state formants. Statistically based pattern classifiers used as models for human perception have shown that very coarse representations of formant change over time result in accurate classification of American English vowels. In this study, using synthetic stimuli with five levels of formant contour detail, human listeners achieved maximum vowel identification for relatively coarse representations of formant movement containing information about onset, offset, and midpoint frequencies. More detailed representations of contour did not improve identification for most vowels.

Time-varying spectral change in the vowels of children and adults.

Recent studies have shown that time-varying changes in formant pattern contribute to the phonetic specification of vowels. This variation could be especially important in children's vowels, because children have higher fundamental frequencies (f0's) than adults, and formant-frequency estimation is generally less reliable when f0 is high. To investigate the contribution of time-varying changes in formant pattern to the identification of children's vowels, three experiments were carried out with natural and synthesized versions of 12 American English vowels spoken by children (ages 7, 5, and 3 years) as well as adult males and females. Experiment 1 showed that (i) vowels generated with a cascade formant synthesizer (with hand-tracked formants) were less accurately identified than natural versions; and (ii) vowels synthesized with steady-state formant frequencies were harder to identify than those which preserved the natural variation in formant pattern over time. The decline in intelligibility was similar across talker groups, and there was no evidence that formant movement plays a greater role in children's vowels compared to adults. Experiment 2 replicated these findings using a semi-automatic formant-tracking algorithm. Experiment 3 showed that the effects of formant movement were the same for vowels synthesized with noise excitation (as in whispered speech) and pulsed excitation (as in voiced speech), although, on average, the whispered vowels were less accurately identified than their voiced counterparts. Taken together, the results indicate that the cues provided by changes in the formant frequencies over time contribute materially to the intelligibility of vowels produced by children and adults, but these time-varying formant frequency cues do not interact with properties of the voicing source.

Perception of linear formant trajectories in vowels

Boundary locations for identification functions were tested for /i-eh/ and /u-i/ continua. In the /i-eh/ continuum, F1 varied in 14 steps from 360 to 580 Hz, and in the /u-i/ continuum, F2 varied in 19 steps from 1000 to 2000 Hz. There were five different types of stimuli in each continuum with (1) steady-state formants, (2) one formant changing linearly over time, and (3) two formants changing linearly. The single changing formant was F1 in the /i–eh/ continuum and F2 in the /u–i/ continuum. The two changing formants were F1 and F2 in both continua. The trajectories had upward or downward directions. Identification functions were obtained with ten normal-hearing subjects. Directions of boundary shifts for the /i–eh/ continuum indicated a perceptual emphasis of the initial segments of F1 and F2 trajectories. The directions of boundary shifts for the /u–i/ continuum indicated a perceptual emphasis of the final segment of F2 trajectory when only F2 changed over time. However, when changes over time in F1 were added, boundaries for these stimuli were not significantly different from the boundary for the stimuli with steady-state formants. Differences in perception of F1 and F2 trajectories in two tested continua will be discussed. [Work supported by NIH.]

A statistical model for formant-transition microsegments of speech incorporating locus equations

Abstract Locus equations, which describe linear relationships between the onset and steady-state formant values in consonant-vowel syllables, have recently been measured using a large quantity of acoustic data and have been proposed as a source of relational invariance for stop place categorization (H. Sussman et al., 1991). In this paper we present a statistical model which utilizes the conceptualization of the locus equations as a basis for parametric of modeling of phonetic contexts — place of articulation, and of their acoustic consequences — formant transitions. The model is based on a hidden Markov model representation of formant-transition microsegments of speech. We develop a generalized EM algorithm for automatic estimation of the model parameters. The proposed model is capable of generalizing consonant characteristics from a small training data set where the contextual information is only sparsely represented, and is hence applicable to solving very large vocabulary speech recognition problems.

Vowel boundaries for steady-state and linear formant trajectories

Locations of boundaries and slopes of identification functions were tested for /I-epsilon/ vowel continua with steady-state and linearly changing formant trajectories. In experiment 1, the boundaries and slopes for arbitrarily selected trajectory directions were determined for ten normal-hearing and ten hearing-impaired subjects in three listening conditions: Quiet, noise, and reverberation. The boundaries did not depend upon the group of subjects or the listening condition. A boundary shift was found for stimuli with F1 changing in a downward direction relative to boundaries for stimuli with either only F1 or with both F1 and F2 changing in an upward direction. The slope of the identification function for stimuli with F1 changing in a downward direction was shallower than the slopes for stimuli with steady-state formants or stimuli with F1 changing in an upward direction. The slopes obtained from the hearing-impaired subjects were shallower than those of the normal-hearing subjects and were shallower in noise than in either quiet or reverberation. In experiment 2, boundaries and slopes for the trajectory directions found in the natural vowels /I/ and /epsilon/, F1 changing in an upward direction and F2 in a downward direction, were determined for nine normal-hearing subjects in two listening conditions, quiet and reverberation. The boundary for stimuli with both F1 and F2 changing in directions characteristic for natural vowels was shifted relative to the boundary for stimuli with steady-state formants. The directions of the boundary shifts in experiments 1 and 2 indicated a perceptual emphasis on the initial sections of changing F1 and F2. Sound quality of the end-point /I/ and /epsilon/ stimuli depended upon F1 and F2 trajectories. For both vowels, the best quality judgments were found for the stimuli with natural F1 and F2 trajectory directions. The quality judgments were weakly correlated with the slopes of identification functions, with better quality judgments being associated with steeper slopes.

Acoustic vowel reduction as a function of sentence accent, word stress, and word class

The effect of sentence accent, word stress, and word class (function words versus content words) on the acoustic properties of 9 Dutch vowels in fluent speech was investigated. A list of sentences was read aloud by 15 male speakers. Each sentence contained one syllable of interest. This could be a monosyllabic function word, an unstressed syllable of a content word, or a stressed syllable of a content word. The same syllable occurred in all three conditions. Sentence accent was manipulated with questions that preceded the sentences. A total number of 3465 vowels were segmented from the syllables and analysed. It was found that all three factors mentioned above had a significant effect both on the steady-state formant frequencies ( F 1 and F 2) and on the duration of the vowels. Word stress and word class had a stronger effect on the vowels than sentence accent. A listening experiment showed the perceptual significance of the acoustic measurements. It appeared that spectral vowel reduction could be better interpreted as the result of an increased contextual assimilation than as the tendency to centralize. We also studied changes in the dynamics of the formant tracks due to the experimental conditions. It was found that formant tracks of reduced vowels became flatter which supports the view of an increased contextual assimilation. Three simple models of vowel reduction are discussed.

Vowel boundaries for steady-state and transient formants.

Fifteen-step /i–eh/ continua were generated to investigate effects of: (1) listening conditions (quiet, noise, and reverberation), (2) subjects’ hearing (normal and impaired), and (3) trajectories of F1 and F2 (all steady-state, F1 changing in upward or downward direction, and both F1 and F2 changing in upward or downward direction). Stimuli (each repeated 10 times) generated with the Klatt synthesizer were 200 ms long. Speech-spectrum noise was mixed at S/N=0 dB. Reverberation was generated by a computer program and convolved with the stimuli (T=0.9 s). Subjects, 10 in each group, were tested individually. Stimuli were delivered monaurally through an earphone at a comfortable level. Listening condition had no effect on the boundary location for either group of subjects and had no effect on boundary slope for the normal-hearing subjects but a significant effect on slope for hearing-impaired subjects. Type of trajectories had a significant effect on location of the boundary only for the hearing-impaired subjects and on boundary slope for both groups of subjects; the two continua with F1 and with F1 and F2 changing in downward direction produced different results than the continua with steady-state formants and with formants in upward direction. [Work supported by NIDCD.]

Tonotopic features of speech-evoked activity in primate auditory cortex

To further clarify the neural mechanisms underlying the cortical encoding of speech sounds, we have recorded multiple unit activity (MUA) in the primary auditory cortex (A1) and thalamocortical (TC) radiations of an awake monkey to 3 consonant-vowel syllables, /da/, /ba/ and /ta/, that vary in their consonant place of articulation and voice onset time (VOT). In addition, we have examined the responses to the syllables' isolated formants and formant pairs. Response features are related to the cortical tonotopic organization, as determined by examining the responses to selected pure tones. MUA patterns that differentially reflect the spectral characteristics of the steady-state formant frequencies and formant transition onset frequencies underlying consonant place of articulation occur at sites with similarly differentiated tone responses. Whereas the detailed spectral characteristics of the speech sounds are reflected in low frequency cortical regions, both low and high frequency areas generate responses that reflect their temporal characteristics of fundamental frequency and VOT. Formant interactions modulate the responses to the whole syllables. These interactions may sharpen response differences that reflect consonant place of articulation. Response features noted in A1 also occur in TC fibers. Thus, differences in the encoding of speech sounds between the thalamic and cortical levels may include further opportunities for formant interactions within auditory cortex. One effect could be to heighten response contrast between complex stimuli with subtle acoustical differences.

Identification of vowels in "vowel-less" syllables by 3-year-olds.

The ability of 3-year-old children to perceive the identity of vowels in full-vowel and silent-center, consonant-vowel-consonant (CVC) syllables was investigated using a two-alternative pointing procedure. Silence replaced the middle 10%, 35%, 65%, or 90% of the steady-state formants of synthetic "bad" and "bud" syllables. Identification of the two full-vowel syllables was 87% correct, whereas performance for the silent-center syllables was somewhat lower (72%, 70%, 67%, and 66% correct for the 10%, 35%, 65%, and 90% deletion stimuli, respectively). The performance of individual children fell into two subgroups: (1) those who performed like adults by maintaining correct vowel identification for all of the silent-center syllables, and (2) those who identified the full-vowel syllables correctly but performed at chance for all of the silent-center syllables. Three additional experiments showed that none of the children performed poorly when noise replaced the gap in the silent-center syllables. These results demonstrate that many 3-year-olds can identify vowels correctly in CVC syllables in the absence of the full spectral properties of steady-state formants.

Representation of stop consonants in the discharge patterns of auditory-nerve fibers

The representation of the speech syllables /da/ and /ba/ in populations of auditory-nerve fibers was studied. Post-stimulus-time histograms were computed from 20-ms segments of fiber spike trains occurring in response to the stimulus. Discrete Fourier transforms with a resolution of 50 Hz were computed from each histogram. As a measure of the response of the population of fibers to each harmonic of the 50-Hz resolution frequency of the transform, the magnitude of the response to that frequency was averaged across all fibers whose characteristic frequencies were within one-fourth octave of that harmonic. We have previously called this measure the average localized synchronized rate (ALSR). Response profiles for the 20-ms segments of the stimulus were generated by plotting the ALSR versus frequency. Time-varying spectral features of the /da/ and /ba/ stimuli are well preserved by such profiles. For example, the onset spectrum and formant transitions of the consonant-vowel syllable are well represented. Furthermore, the fine structure in the speech spectrum related to the pitch of the excitation source is maintained in these ALSR plots. Average discharge rate profiles were generated in a manner similar to that for the ALSR; in this case average rate replaces Fourier transform components as response measure. Such average rate profiles can represent the transitions of at least formants two and three. However, such average rate profiles do not represent the steady-state formants or the voice pitch.

Synthesis of Final /z/ without Voicing

American English often shows a tendency toward devoicing of “voiced” consonants, particularly in final position, without however losing the “voiced-voiceless” distinction. What simple pattern variables on a speech synthesizer, it was asked, would elicit consistent listener recognition of a “voiced” consonant without recourse to voicing? In an experimental study of this question a small group of listeners identified as “peas,” “peace,” “peed,” “peat,” or “pea,” a set of synthetic speech patterns in which the following features were varied: (1) the duration of a high-frequency final segment representing alveolar friction; (2) the duration of the total pattern (achieved by modifying the length of the steady-state formants representing the vowel); (3) point of buzz to hiss changeover; and (4) the presence or absence of an intensity drop at the end of the vowel formants. All the variables did in fact influence listener identification of the patterns in varying degrees and combinations. In particular, a number of patterns were consistently identified as “peas,” i.e., the word ending with “voiceless” /z/. There were noticeable differences among individual listeners in their sensitivity to the variables and in the ranges of maximum effectiveness. [This research was supported by the Department of Defense in connection with Contract DA 18-119-sc-260.]