A study of acoustic-to-articulatory inversion of speech by analysis-by-synthesis using chain matrices and the Maeda articulatory model

In this paper, a quantitative study of acoustic-to-articulatory inversion for vowel speech sounds by analysis-by-synthesis using the Maeda articulatory model is performed. For chain matrix calculation of vocal tract (VT) acoustics, the chain matrix derivatives with respect to area function are calculated and used in a quasi-Newton method for optimizing articulatory trajectories. The cost function includes a distance measure between natural and synthesized first three formants, and parameter regularization and continuity terms. Calibration of the Maeda model to two speakers, one male and one female, from the University of Wisconsin x-ray microbeam (XRMB) database, using a cost function, is discussed. Model adaptation includes scaling the overall VT and the pharyngeal region and modifying the outer VT outline using measured palate and pharyngeal traces. The inversion optimization is initialized by a fast search of an articulatory codebook, which was pruned using XRMB data to improve inversion results. Good agreement between estimated midsagittal VT outlines and measured XRMB tongue pellet positions was achieved for several vowels and diphthongs for the male speaker, with average pellet-VT outline distances around 0.15 cm, smooth articulatory trajectories, and less than 1% average error in the first three formants.     

Estimating the control parameters of an articulatory model from electromagnetic articulograph data

Finding the control parameters of an articulatory model that result in given acoustics is an important problem in speech research. However, one should also be able to derive the same parameters from measured articulatory data. In this paper, a method to estimate the control parameters of the the model by Maeda from electromagnetic articulography (EMA) data, which allows the derivation of full sagittal vocal tract slices from sparse flesh-point information, is presented. First, the articulatory grid system involved in the model's definition is adapted to the speaker involved in the experiment, and EMA data are registered to it automatically. Then, articulatory variables that correspond to measurements defined by Maeda on the grid are extracted. An initial solution for the articulatory control parameters is found by a least-squares method, under constraints ensuring vocal tract shape naturalness. Dynamic smoothness of the parameter trajectories is then imposed by a variational regularization method. Generated vocal tract slices for vowels are compared with slices appearing in magnetic resonance images of the same speaker or found in the literature. Formants synthesized on the basis of these generated slices are adequately close to those tracked in real speech recorded concurrently with EMA.     

The relationship of vocal tract shape to three voice qualities

Three-dimensional vocal tract shapes and consequent area functions representing the vowels [i, ae, a, u] have been obtained from one male and one female speaker using magnetic resonance imaging (MRI). The two speakers were trained vocal performers and both were adept at manipulation of vocal tract shape to alter voice quality. Each vowel was performed three times, each with one of the three voice qualities: normal, yawny, and twangy. The purpose of the study was to determine some ways in which the vocal tract shape can be manipulated to alter voice quality while retaining a desired phonetic quality. To summarize any overall tract shaping tendencies mean area functions were subsequently computed across the four vowels produced within each specific voice quality. Relative to normal speech, both the vowel area functions and mean area functions showed, in general, that the oral cavity is widened and tract length increased for the yawny productions. The twangy vowels were characterized by shortened tract length, widened lip opening, and a slightly constricted oral cavity. The resulting acoustic characteristics of these articulatory alterations consisted of the first two formants (F1 and F2) being close together for all yawny vowels and far apart for all the twangy vowels.     

Articulatory tradeoffs reduce acoustic variability during American English /r/ production.

The American English phoneme /r/ has long been associated with large amounts of articulatory variability during production. This paper investigates the hypothesis that the articulatory variations used by a speaker to produce /r/ in different contexts exhibit systematic tradeoffs, or articulatory trading relations, that act to maintain a relatively stable acoustic signal despite the large variations in vocal tract shape. Acoustic and articulatory recordings were collected from seven speakers producing /r/ in five phonetic contexts. For every speaker, the different articulator configurations used to produce /r/ in the different phonetic contexts showed systematic tradeoffs, as evidenced by significant correlations between the positions of transducers mounted on the tongue. Analysis of acoustic and articulatory variabilities revealed that these tradeoffs act to reduce acoustic variability, thus allowing relatively large contextual variations in vocal tract shape for /r/ without seriously degrading the primary acoustic cue. Furthermore, some subjects appeared to use completely different articulatory gestures to produce /r/ in different phonetic contexts. When viewed in light of current models of speech movement control, these results appear to favor models that utilize an acoustic or auditory target for each phoneme over models that utilize a vocal tract shape target for each phoneme.     

Adaptive computation of articulatory parameters from the speech signal

An unconstrained optimization technique is used to find the values of parameters, of a combination of an articulatory and a vocal tract model, that minimize the difference between model spectra and natural speech spectra. The articulatory model is anatomically realistic and the vocal tract model is a "lossy" Webster equation for which a method of solution is given. For English vowels in the steady state, anatomically reasonable articulatory configurations whose corresponding spectra match those of human speech to within 2 dB have been computed in fewer than ten iterations. Results are also given which demonstrate a limited ability of the system to track the articulatory dynamics of voiced speech.     

Automatic speech recognition using articulatory features from subject-independent acoustic-to-articulatory inversion

An automatic speech recognition approach is presented which uses articulatory features estimated by a subject-independent acoustic-to-articulatory inversion. The inversion allows estimation of articulatory features from any talker's speech acoustics using only an exemplary subject's articulatory-to-acoustic map. Results are reported on a broad class phonetic classification experiment on speech from English talkers using data from three distinct English talkers as exemplars for inversion. Results indicate that the inclusion of the articulatory information improves classification accuracy but the improvement is more significant when the speaking style of the exemplar and the talker are matched compared to when they are mismatched.     

Evolving theories of vowel perception

Research on the perception of vowels in the last several years has given rise to new conceptions of vowels as articulatory, acoustic, and perceptual events. Starting from a "simple" target model in which vowels were characterized articulatorily as static vocal tract shapes and acoustically as points in a first and second formant (F1/F2) vowel space, this paper briefly traces the evolution of vowel theory in the 1970s and 1980s in two directions. (1) Elaborated target models represent vowels as target zones in perceptual spaces whose dimensions are specified as formant ratios. These models have been developed primarily to account for perceivers' solution of the "speaker normalization" problem. (2) Dynamic specification models emphasize the importance of formant trajectory patterns in specifying vowel identity. These models deal primarily with the problem of "target undershoot" associated with the coarticulation of vowels with consonants in natural speech and with the issue of "vowel-inherent spectral change" or diphthongization of English vowels. Perceptual studies are summarized that motivate these theoretical developments.     

Variability in production of the vowels /i/ and /a/

A hypothesis on the nature of articulatory targets for the vowels /i/ and /a/ is proposed, based on acoustic considerations and vowel articulations. The conjecture is that positioning of points on the tongue surface in a repetition experiment should be most accurate in the direction perpendicular to the vocal-tract midline, at the acoustically critical point of maximal constriction for each vowel. The hypothesis was tested by: examining x-ray microbeam data for three speakers, conducting a partial acoustical analysis, and performing a modeling study. Distributions were plotted of the midsagittal locations of three tongue points at the time of maximal excursion toward the vowel target for numbers of examples of the vowels, embedded in a variety of phonetic contexts. More variation was found along a direction parallel to the vocal tract midline than perpendicular to the midline, supporting the hypothesis. Statistics on formant values for one subject have been calculated, and pairwise regressions of displacement and formant data have been run. An articulatory synthesizer [Rubin et al., J. Acoust. Soc. Am. 70, 321-328 (1981)] has been manipulated through displacements similar to the subject's articulatory variation. Although articulatory synthesis showed systematic relationships between articulatory relationships and formant frequencies, there were no significant correlations between the subject's measured articulatory displacements and his formant data. These additional results raise questions about the methodology and point to the need for additional work for an adequate test of the hypothesis.     

Vocal tract normalization for midsagittal articulatory recovery with analysis-by-synthesis.

A method is presented that accounts for differences in the acoustics of vowel production caused by human talkers' vocal-tract anatomies and postural settings. Such a method is needed by an analysis-by-synthesis procedure designed to recover midsagittal articulatory movement from speech acoustics because the procedure employs an articulatory model as an internal model. The normalization procedure involves the adjustment of parameters of the articulatory model that are not of interest for the midsagittal movement recovery procedure. These parameters are adjusted so that acoustic signals produced by the human and the articulatory model match as closely as possible over an initial set of pairs of corresponding human and model midsagittal shapes. Further, these initial midsagittal shape correspondence need to be generalized so that all midsagittal shapes of the human can be obtained from midsagittal shapes of the model. Once these procedures are complete, the midsagittal articulatory movement recovery algorithm can be used to derive model articulatory trajectories that, subsequently, can be transformed into human articulatory trajectories. In this paper the proposed normalization procedure is outlined and the results of experiments with data from two talkers contained in the X-ray Microbeam Speech Production Database are presented. It was found to be possible to characterize these vocal tracts during vowel production with the proposed procedure and to generalize the initial midsagittal correspondences over a set of vowels to other vowels. The procedure was also found to aid in midsagittal articulatory movement recovery from speech acoustics in a vowel-to-vowel production for the two subjects.     

Acoustic and perceptual effects of changes in vocal tract constrictions for vowels.

The purpose of this study was to use vocal tract simulation and synthesis as means to determine the acoustic and perceptual effects of changing both the cross-sectional area and location of vocal tract constrictions for six different vowels: Area functions at and near vocal tract constrictions are considered critical to the acoustic output and are also the central point of hypotheses concerning speech targets. Area functions for the six vowels, [symbol: see text] were perturbed by changing the cross-sectional area of the constriction (Ac) and the location of the constriction (Xc). Perturbations for Ac were performed for different values of Xc, producing several series of acoustic continua for the different vowels. Acoustic simulations for the different area functions were made using a frequency domain model of the vocal tract. Each simulated vowel was then synthesized as a 1-s duration steady-state segment. The phoneme boundaries of the perturbed synthesized vowels were determined by formal perception tests. Results of the perturbation analyses showed that formants for each of the vowels were more sensitive to changes in constriction cross-sectional area than changes in constriction location. Vowel perception, however, was highly resistant to both types of changes. Results are discussed in terms of articulatory precision and constriction-related speech production strategies.     

The magnetic resonance imaging subset of the mngu0 articulatory corpus

This paper announces the availability of the magnetic resonance imaging (MRI) subset of the mngu0 corpus, a collection of articulatory speech data from one speaker containing different modalities. This subset comprises volumetric MRI scans of the speaker's vocal tract during sustained production of vowels and consonants, as well as dynamic mid-sagittal scans of repetitive consonant-vowel (CV) syllable production. For reference, high-quality acoustic recordings of the speech material are also available. The raw data are made freely available for research purposes.     

A study of high front vowels with articulatory data and acoustic simulations

The purpose of this study is to test a methodology for describing the articulation of vowels. High front vowels are a test case because some theories suggest that high front vowels have little cross-linguistic variation. Acoustic studies appear to show counterexamples to these predictions, but purely acoustic studies are difficult to interpret because of the many-to-one relation between articulation and acoustics. In this study, vocal tract dimensions, including constriction degree and position, are measured from cinéradiographic and x-ray data on high front vowels from three different languages (North American English, French, and Mandarin Chinese). Statistical comparisons find several significant articulatory differences between North American English /i/ and Mandarin Chinese and French /i/. In particular, differences in constriction degree were found, but not constriction position. Articulatory synthesis is used to model the acoustic consequences of some of the significant articulatory differences, finding that the articulatory differences may have the acoustic consequences of making the latter languages' /i/ perceptually sharper by shifting the frequencies of F(2) and F(3) upwards. In addition, the vowel /y/ has specific articulations that differ from those for /i/, including a wider tongue constriction, and substantially different acoustic sensitivity functions for F(2) and F(3).     

Time dependence of vocal tract modes during production of vowels and vowel sequences

Vocal tract shaping patterns based on articulatory fleshpoint data from four speakers in the University of Wisconsin x-ray microbeam (XRMB) database [J. Westbury, UW-Madison, (1994)] were determined with a principal component analysis (PCA). Midsagittal cross-distance functions representative of approximately the front 6 cm of the oral cavity for each of 11 vowels and vowel-vowel (VV) sequences were obtained from the pellet positions and the hard palate profile for the four speakers. A PCA was independently performed on each speaker's set of cross-distance functions representing static vowels only, and again with time-dependent cross-distance functions representing vowels and VV sequences. In all cases, results indicated that the first two orthogonal components (referred to as modes) accounted for more than 97% of the variance in each speaker's set of cross-distance functions. In addition, the shape of each mode was shown to be similar across the speakers suggesting that the modes represent common patterns of vocal tract deformation. Plots of the resulting time-dependent coefficient records showed that the four speakers activated each mode similarly during production of the vowel sequences. Finally, a procedure was described for using the time-dependent mode coefficients obtained from the XRMB data as input for an area function model of the vocal tract.     

A generalized smoothness criterion for acoustic-to-articulatory inversion

The many-to-one mapping from representations in the speech articulatory space to acoustic space renders the associated acoustic-to-articulatory inverse mapping non-unique. Among various techniques, imposing smoothness constraints on the articulator trajectories is one of the common approaches to handle the non-uniqueness in the acoustic-to-articulatory inversion problem. This is because, articulators typically move smoothly during speech production. A standard smoothness constraint is to minimize the energy of the difference of the articulatory position sequence so that the articulator trajectory is smooth and low-pass in nature. Such a fixed definition of smoothness is not always realistic or adequate for all articulators because different articulators have different degrees of smoothness. In this paper, an optimization formulation is proposed for the inversion problem, which includes a generalized smoothness criterion. Under such generalized smoothness settings, the smoothness parameter can be chosen depending on the specific articulator in a data-driven fashion. In addition, this formulation allows estimation of articulatory positions recursively over time without any loss in performance. Experiments with the MOCHA TIMIT database show that the estimated articulator trajectories obtained using such a generalized smoothness criterion have lower RMS error and higher correlation with the actual measured trajectories compared to those obtained using a fixed smoothness constraint.     

The synergy between speech production and perception

Speech intelligibility is known to be relatively unaffected by certain deformations of the acoustic spectrum. These include translations, stretching or contracting dilations, and shearing of the spectrum (represented along the logarithmic frequency axis). It is argued here that such robustness reflects a synergy between vocal production and auditory perception. Thus, on the one hand, it is shown that these spectral distortions are produced by common and unavoidable variations among different speakers pertaining to the length, cross-sectional profile, and losses of their vocal tracts. On the other hand, it is argued that these spectral changes leave the auditory cortical representation of the spectrum largely unchanged except for translations along one of its representational axes. These assertions are supported by analyses of production and perception models. On the production side, a simplified sinusoidal model of the vocal tract is developed which analytically relates a few "articulatory" parameters, such as the extent and location of the vocal tract constriction, to the spectral peaks of the acoustic spectra synthesized from it. The model is evaluated by comparing the identification of synthesized sustained vowels to labeled natural vowels extracted from the TIMIT corpus. On the perception side a "multiscale" model of sound processing is utilized to elucidate the effects of the deformations on the representation of the acoustic spectrum in the primary auditory cortex. Finally, the implications of these results for the perception of generally identifiable classes of sound sources beyond the specific case of speech and the vocal tract are discussed.     

Articulatory dynamics of loud and normal speech

A comparison was made between normal and loud productions of bilabial stops and stressed vowels. Simultaneous recordings of lip and jaw movement and the accompanying audio signal were made for four native speakers of Swedish. The stimuli consisted of 12 Swedish vowels appearing in an /i'b_b/ frame and were produced with both normal and increased vocal effort. The displacement, velocity, and relative timing associated with the individual articulators as well as their coarticulatory interactions were studied together with changes in acoustic segmental duration. It is shown that the production of loud as compared with normal speech is characterized by amplification of normal movement patterns that are predictable for the above articulatory parameters. In addition, it was observed that the acoustic durations of bilabial stops were shortened, whereas stressed vowels were lengthened during loud speech production. Two interpretations of the data are offered, viewing loud articulatory behavior as a response to production demands and perceptual constraints, respectively.     

Measurement of temporal changes in vocal tract area function from 3D cine-MRI data

A 3D cine-MRI technique was developed based on a synchronized sampling method [Masaki et al., J. Acoust. Soc. Jpn. E 20, 375-379 (1999)] to measure the temporal changes in the vocal tract area function during a short utterance /aiueo/ in Japanese. A time series of head-neck volumes was obtained after 640 repetitions of the utterance produced by a male speaker, from which area functions were extracted frame-by-frame. A region-based analysis showed that the volumes of the front and back cavities tend to change reciprocally and that the areas near the larynx and posterior edge of the hard palate were almost constant throughout the utterance. The lower four formants were calculated from all the area functions and compared with those of natural speech sounds. The mean absolute percent error between calculated and measured formants among all the frames was 4.5%. The comparison of vocal tract shapes for the five vowels with those from the static MRI method suggested a problem of MRI observation of the vocal tract: data from static MRI tend to result in a deviation from natural vocal tract geometry because of the gravity effect.     

Articulatory correlates of stress and speaking rate in Swedish VCV utterances

Articulatory activity underlying changes in stress and speaking rate was studied by means of x-ray cinefilm and acoustic speech records. Two Swedish subjects produced vowel-consonant-vowel (VCV) utterances under controlled rate-stress conditions. The vowels were tense (i a u), and the consonants were the voiceless stops, notably (p). The spectral characteristics of the vowels were not significantly influenced by changes in the speaking rate. They were, however, significantly emphasized under stress. At the articulatory level, stressed vowels displayed narrower oral tract constrictions than unstressed vowels at the two speaking rates studied. At the faster speaking rate, vowel- and consonant-related gestures were coproduced to a greater extent than at the slower rate. The data, failing to produce evidence for an "undershoot" mechanism, support the view that dialect-specific correlates of stress are actively safeguarded by means of articulatory reorganization.