Speech Commun Research Articles

Spectral slope as a cue for the perception of breathy and non-breathy stops in Shanghainese

In Shanghainese, breathy excitation following release distinguishes category 2 (so-called “voiced unaspirated”) stops from category 1 (“voiceless unaspirated”) stops. The most obvious acoustic correlate of bresthiness is relative spectral slope: During the first 50–100 ms after release, H1-H2 is greater for category 2 than for category 1 in both word-initial and word-medial (morpheme-initial) stops (Ren, unpublished). To investigate the perception of breathiness, a series of synthetic disyllables was prepared in which the spectral slope during the 100 ms following the release of the word-medial stop was increased step by step increasing the value of the “open quotient” in the computation of the glottal waveform—a technique previously used for the synthesis of breathy vowels [C. Bickley, Work. Papers MIT Speech Commun. Lab. 1, 71–81 (1982)]. Other potential cues (closure duration, pulsing during closure, tone contour) were neutralized. Ten Shanghainese speakers labeled the stimuli with greater spectral slope as category 2 and those with smaller spectral slope as category 1. [Work supported by NICHD Grant 01994.]

Evidence for the 3‐Bark integration interval

It is generally accepted that sound energy within a 1‐Bark interval is integrated by the ear. Chistovich et al. [in Frontiers of Speech Communication Research, edited by Lindblom and Ohman (Academic, New York, 1979), pp. 143–157] found evidence of a larger integration interval, of approximately 3 Bark, for vowel perception. Klatt [J. Acoust. Soc. Am. Suppl. 1 77, S7 (1985)] found that listeners could distinguish between vowels whose formant differences were compensated by bandwidth differences (so that cue trading did not occur) and concluded that the 3‐Bark interval was not supported. The present paper analyzes the effect of different intervals on the center of gravity analysis described in Jaykin et al. [J. Acoust. Soc. Am. Suppl. 1 82, S81 (1988)] adapted from Chistovich and Chernova [Speech Commun. 5, 3–16 (1986)], and compares that analysis with a perceptual experiment testing the effect of harmonics on the perception of formants. The results support the hypothesis that the minimum interval is about 3 Bark. Our model may explain the lack of cue trading between formants and bandwidths.

Voicing in French and English labial stops: A cross‐language perceptual study

Measurements of European French and Canadian English labial stops indicate that the French /p/ and English /b/ categories are very similar, often exhibiting only small differences in voice onset time. Two groups of English‐speaking listeners (one with some knowledge of French, the other without) were asked to categorize a set of modified natural tokens from these categories as either /p/ or /b/. The tokens were taken from word‐initial stops produced in a sentence context. Components preceding the release bursts were removed, and the signals were truncated at 68 ms. Despite the small difference in VOT between the categories, most listeners were able to reliably separate them at levels above chance. Analysis revealed that the listeners may have relied primarily on VOT in making their judgments and that overall amplitude [C. J. Darwin and M. Pearson, Speech Commun. 1, 29–44 (1982)] may have played a secondary role. These findings indicate that listeners may be sensitive to small differences between categories in their native language and analogous categories in a foreign language.

Coding of articulatory‐acoustic events in a model of the peripheral auditory system

In the last years, several investigators have been concerned with speech processing in auditory models. These lines of research were carefully followed and research was restricted to auditory detection of a set of articulatory and acoustic events. The strategy consists of exploring various processing abilities lying in the model of the peripheral auditory system developed (see Dolmazon and Boulogne, Speech Commun. 1, 55–73), and attempting to demonstrate a possible adequacy between some of the physiological abilities and the acoustical structure of speech induced by articulatory constraints. Here, first results about characterization of voicing onsets and offsets and frication onsets and offsets for CVCV utterances in French will be presented and results obtained by various combinations of physiological mechanisms will be compared: with and without neural adaptation, with enhancement or diminution of low‐frequency characteristics by temporally or geographically based filtering processes.

Voiced speech model including source‐tract interaction

This paper considers the problem of estimating the parameters of a voiced speech model that includes the effects of source‐tract interaction. Based on work by Ananthapadmanabha and Fant [Speech Commun. 1, 167–184 (1982)], the voiced speech signal is modeled as the output of a time‐varying vocal tract filter excited by a parametrized glottal source waveform. The glottal source waveform [Fant et al., STL/QPSR (1984)] models the main pulse shape of the glottal volume velocity and is described by four parameters. The effect of source‐tract interaction is modeled by varying the frequency and bandwidth of the first formant in synchrony with the glottal source waveform. An analysis‐by‐synthesis approach is used to estimate the parameters of the model. Algorithms for estimating the parameters of the glottal source waveform and the vocal tract filter are described. Comparisons of the spectrum of the original and synthesized speech waveforms are presented. [Work supported by the OSU Seed grant program.]

The “trough” phenomenon in Turkish and in English

Previous studies on several Indo‐European languages have shown that for strings of two rounded vowels separated by phonologically rounding‐neutral consonants, both EMG and lip protrusion movement traces show double peaks coincident with the two rounded vowels plus an intervening “trough” [Bell‐Berti and Harris, J. Acoust. Soc. Am. 71, 449–454 (1982); Perkell, Speech Commun. 5, 47–68]. This has been interpreted as due to independent rounding gestures associated with each vowel and as counterevidence to the claim that rounding coarticulation is anticipatory. In this paper, English speakers' production of nonsense words /kuktluk/, /kuktuk/, /kukuk/, /kutuk/, and /kuluk/ will be compared with that of speakers of Turkish, a language with strong constraints against the cooccurence of rounded and unfounded vowels in the same word. Both movement and EMG evidence will be used to argue that the trough phenomenon is more extensive in English than in Turkish. Implications for cross‐language differences in coarticulat...

Masking period patterns: presentation of a model consistent with an enlarged set of experimental results

Masking period patterns (MPP) have been proposed by Zwieker as psychoaoustic equivalent of period histograms (PSTH) measured in auditory neurons. A large set of results have been obtained and represented on a model for predicting MPP [E. Zwicker, Biol. Cybern. 23, 49–60 (1976)]. We have already used this technique for studying PSTH when the basilar membrane is stimulated by trapezoidal patterns of excitation [P. Escudier et al., J. Acoust. Soc. Am. Suppl. 1 67, S102 (1980)]. Our results cannot be completely interpreted on the basis of Zwicker's model. We present a new model, using a peripheral auditory system simulation previously elaborated in our group [J. M. Dolmazon et al., Speech Commun. 1, 52–73 (1982)] and some simple additional hypotheses. This model is shown to be able to predict the whole set of available MPP results.

Computation of vocal tract sensitivity functions: A complex Lagrangian approach

A method for computing vocal tract sensitivity functions is presented in this paper. Fant and Pauli [Speech Commun. Semin., Stockolm, 121–132 (1974)] have derived closed‐form expressions relating the kinetic and potential energies in the tract to the sensitivity of formant frequencies with respect to small variations of the area function. In our approach, more general closed‐form formulas are derived to compute the sensitivity of any quantity related to the vocal tract transfer function. A Lagrangian formulation allows a straightforward computation of the gradient of acoustic parameters with respect to area function parameters. This technique was used to replace the finite‐difference gradient approximation in a procedure identifying articulatory parameters from formant parameters [Charpentier, ICASSP Proc., 1984–1987 (1982)]. The method takes into account the vocal tract losses (heat conduction, viscosity, wall vibrations, lip radiation). It was found effective for determining the articulatory parameters from outputs of the model itself.

Frequency analyses of American English vowels

Analyses were made of acoustic data derived from spectrographic measurements of ten American English vowels spoken by 15 children, 33 men, and 28 women talkers [G. E. Peterson and H. L. Barney, J. Acoust. Soc. Am. 24, 175–184 (1952)]. Frequencies of F0, F1, F2, and F3 were converted to the Bark scale, an auditory scale of perceived pitch in units of critical bands. A critical distance between formants of about 3.5 Bark has been reported to be important in the perception of synthetic Russian vowel quality [L. A. Chistovich, R. L. Sheikin, and V. V. Lublinskaya, Frontiers Speech Commun. Res., 143–157 (1979)]. Three distance measures were calculated from the Bark‐transformed Peterson and Barney data: F1−F0, F2−F1, and F3−F2. Each of the three distance measures was binarily classified according to whether or not frequency differences exceeded 3.5 Bark. These binary classifications related to the linguistic dimensions of vowel height, compactness, and place of articulation. The procedure also served as a normalization technique. [Work supported by NIH and performed in part at the Research Lab. of Electronics, MIT.]

Lateral excitation of a vocal tract analog

A schematic mechanical analog of the vocal tract was constructed, consisting of a cascade of 21 adjustable‐width rectangular cross‐sections of 8‐mm length each. Side openings spaced at 25.4‐mm intervals along the tube permitted the excitation of the tract by a standard hearing aid driver, mounted flush with the internal wall, at any one of seven locations from “glottis” to “lips.” The tract was adjusted to simulate the area functions of several American vowel sounds estimated directly from the speech wave [H. Wakita, Speech Commun. Res. Lab. Monogr. No. 9. (1972)]. The periodic excitation waveform resembled a glottal pulse train; the radiated sound was picked up by a condenser microphone in front of the “lip opening.” Waveshapes and spectra were recorded and formant characteristics measured. Fairly good agreement with the norm was noted in the frequency region up to about 800 Hz. Deviations at higher frequencies can be attributed to the violation of the plane wave propagation usually assumed for the acoustic tube model. This work is part of a collaborative effort (with Dr. Robert J. Hanson of the UCSB Speech Department) to develop a speech support system for laryngectomees. [Project supported by the American Cancer Society, Inc.]

Perceptual basis of vowel‐duration differences

Lisker [Glossa, 8(2), 233–246 (1974)] has shown that there existed no viable explanation of vowel duration differences associated with the voicing state of the following consonant. Lisker's conclusion, that such differences are due to constraints of English, does not adequately explain such differences in other languages. Scully [Stockholm Speech Commun. Sem. 327–234 (1974)] suggests that such differences are due to the different aerodynamic conditions that occur depending on the voicing state of the following consonant. Scully's aerodynamic model does not provide resulting entirely consistent with speech data, however, and is therefore open to question. A different kind of explanation may be needed. Experimental evidence from 12 English speakers suggests that voiced (voiceless) consonants make the preceding vowel appear to be longer (shorter). Although these results have yet to be tested with speakers of other languages, they support an interesting hypothesis: the voicing of the following consonant, by continuing the voicing cue present in the vowel, may create the illusion of greater length. Such an illusion, acting as an input to phonological acquisition, may explain the observed vowel duration differences. [Work supported by NSF.]

WINDBAG—A Vocal-Tract Analog Speech Synthesizer

This report describes the beginnings of an effort in speech synthesis at Brigham Young University. Both terminal analog and vocal tract analog synthesizers will be studied eventually, but this paper discusses only a vocal tract synthesizer. The synthesizer is patterned after one described by Flanagan and Landgraf [IEEE Trans. Audio and Electroacoust. Au-16, 57 (1968)] in which the vocal cords and vocal tract are able to interact. A delay-line scheme similar to one described by Kelly and Lochbaum [Proc. Stockholm Speech Commun. Seminar, R.I.T. Stockholm, Sept. 1962)], but terminated at the mouth end with a filter, is used to represent the vocal tract. Some modifications to Flanagan's vocal-cord model have been incorporated. Preliminary synthesis results for steady-state vowels and consonant-vowel transitions will be presented.