Articulatory Model Research Articles

Package of computer programs for articulatory modeling of the speech process

At the poster session we will describe a package of FORTRAN. programs which accomplish articulatory synthesis of speech from phonetic input. The package of subroutines will include the following: (1) a coarticulation program which, knowing featurelike descriptions of speech segments, performs the staggered sequencing of “motor commands” necessary for proper overlapping of articulatory movements; (2) a dynamics program that converts stepwise motor commands into smooth articulatory movements; (3) an instantaneous model of the vocal tract; (4) a parametric-speed model of the pneumatic system, of vocal-cord activity and fricative noise sources; and (5) a routine for converting the above articulatory details into formants and other controls for parametric synthesis. (Of course, the package of programs can also be used with a direct time-domain acoustic tube model, as well.) The formal presentation will contain a brief description of the package of programs, and comments on several aspects of speech synthesis by articulatory modeling.

Report of the Symposium on Articulatory Modeling, Grenoble, July 11〜12, 1977

Report of the Symposium on Articulatory Modeling, Grenoble, July 11〜12, 1977

Normalization and Adaptation of Articulatory Model

Normalization and Adaptation of Articulatory Model

Numerical inversion of the transformation from articulatory to acoustic parameters in the vocal tract

A method is presented for determining the parameters of an articulatory model corresponding to a given set of formant frequencies. A four‐parameter articulatory model of the vocal tract was used to define the vocal‐tract shape. The articulatory parameters were (a) location of constriction, (b) area of constriction, (c) area of mouth opening, and (d) length of the vocal tract. The model included losses due to finite wall impedance, glottal leakage, friction, heat conduction and radiation. Frequencies, amplitudes, and bandwidths of five formants were computed for each of 30 720 vocal‐tract shapes obtained by uniform sampling along each of the four articulatory dimensions. The data was sorted according to the first three formant frequencies. The articulator positions yielding given values of these formant frequencies can be looked up in the sorted data. One‐dimensional fibers (curved lines) were computed along which the articulator positions can vary without producing changes in the three formant frequencies. These fibers span large distances in the articulatory space. Sounds produced by various articulations along a fiber were computed and were compared by informal listening tests.

Preparation of Articulatory Models for Instruction and Display

Preparation of Articulatory Models for Instruction and Display

Articulatory characteristics as a cue for the perception of bilabials /b, w, m/

This paper describes some findings about the role of articulatory parameters in perceiving bilabials /b/, /w/, and /m/ in an intervocalic context (/aCa/), based on speech sounds synthesized with an articulatory model. In the model, the time-varying patterns of lip-opening area were approximated by parabolic functions based on measurements obtained from a high-speed motion picture of the utterances /awa/, /aba/, and /ama/. The parameters specifying the patterns were the minimum value of the lip-opening area (Sm) and the transition duration between the minimum and stationary values of the lip-opening area (Tt). Sm's and Tt's were varied from −0.3 cm2 to 1.0 cm2 (10 items) and from 40 msec to 200 msec (nine items), respectively. The results of the listening tests showed that the distinction between /m/ and /w/ as well as that between /b/ and /w/ depended almost only on the value of Sm, but not on the transition duration Tt. Thus it can be concluded that the degree of lip closure plays a very important role as a cue for distinguishing bilabial consonants in an /aCa/ context, irrespective of the nasalization of the sounds.

Determining tongue body motion from the acoustic speechwave

A technique is described for estimating vocal-tract areas and articulatory parameters from speech sounds. Voiced sounds are analyzed by the covariance method to obtain predictor coefficients. Three or more resonance frequencies are determined; the bandwidth estimates are discarded. Corrections are made to account for inductive yielding walls of the vocal tract. Using a gradient technique, parameters of an articulatory model are manipulated to obtain best fits to the analyzed data. This method avoids the usual tenuous assumptions about vocal-cord spectra and about the distribution of losses in the vocal tract. Results of applying the method to vowels and short words with velar consonants are presented.

An investigation of categorical speech discrimination by rhesus monkeys

The categorical discrimination of synthetic human speech sounds by rhesus macaques was examined using the cardiac component of the orienting response. A within-category change which consisted of stimuli differing acoustically in the onset of F2 and F3 transitions, but which are identified by humans as belonging to thesame phonetic category, were responded to differently from a no-change control condition. Stimuli which differed by the same amount in the onset of F2 and F3 transitions, but which human observers identify as belonging toseparate phonetic categories, were differentiated to an even greater degree than the within-category stimuli. The results provide ambiguous data for an articulatory model of human speech perception and are interpreted instead in terms of a feature-detector model of auditory perception.

Discrimination of vowels and semivowels synthesized with an articulatory model

In order to examine the difference between perceptual manner of vowels and semivowels which are characterized by dynamic acoustic events, an articulatory model for producing both high vowels such as [i] and [w] and the syllable [ju] was first introduced. The formant frequency movements of the syllables obtained from the model were considered to be reasonable as compared with those of natural speech sounds. The difference limens (DLs) of vowels and semivowels for the changes in an articulatory point in the model were measured by an AX discrimination test. The DLs were almost the same among four semivowels having different terminal articulation points when the initial articulation points and lip opening area were kept constant. But the DLs were significantly different among four sustained vowels having the same values as the vocal tract area functions of the semivowels at a terminal point. The implications of these results for feature detecting mechanisms in the perception of vowels and semivowels are discussed with the role of phonetic framework of listeners.

“Quasiphysiological” tongue model

A quasiphysiological dynamic model of the tongue has been devised and implemented in a discrete simulation. The model changes shape only in the saggital plane. Its thickness in the dimension perpendicular to this plane is fixed; consequently, areas in the midsaggital plane correspond to volumes of the tongue. The model consists of 16 mass-bearing fleshpoints, which are connected to one another and to bony attachments by 38 active muscle elements and 47 passive connective-tissue elements. A simple submodel accounts for the behavior of each element type, and the muscle model develops tension in response to an input stimulation. The tongue muscles are represented by anatomical groupings of these tension-developing elements which are arranged to enclose 14 quadrilateral areas. These “volumes” have limited compressibility to allow for the transmission of compression forces and conservation of volume. In addition to tensile and volume-conserving forces, sliding-friction and hard-structure impenetrability forces act on the fleshpoints in the model to determine their movement. The model's behavior is compared with EMG and cineradiographic data, and it will become a component of an overall articulatory model which is intended to guide experimentation and evaluation of data on the physical and physiological properties of the vocal tract. [Research supported by an NIH grant.]

Is the Relationship between Acoustic and Articulatory Variables in Speech Intrinsically Linear?

The functional relationship between the acoustic parameters of speech and the positions of different articulators was investigated on a computer-simulated model of the vocal tract using nonlinear regression. The area function of the vocal tract in the model was specified by seven articulatory variables, namely, the area and location of the maximum constriction formed by the tongue body, the area and location of the tongue tip constriction, the area of the mouth opening, lip protrusion, and the length of the vocal tract. It was found that any one of the seven articulatory variables could be represented reasonably accurately by a polynomial function of the acoustic parameters of the vocal tract model. Two interesting properties of this relationship were observed. First, in estimating an artic-ulatory variable, knowledge of the other articulatory variables was not required; the articulatory variable was uniquely determined by the acoustic parameters alone. Second, any one of several different acoustic representations—such as predictor coefficients, formant frequencies, and their band widths, or power spectrum—could be used with equal success in determining articulatory variables. The applications of this method for determining positions of the articulators during speech production, for speech recognition, and for control of articulatory models for speech synthesis are discussed.

Coarticulation of upper lip protrusion in French.

Abstract This study investigates the temporal organization of upper lip protrusion in French. A photocell was used to transduce upper lip movement into a recordable electrical signal. The utterances under study were recorded by six native speakers of French. Three aspects of the protrusion gesture are examined: extent of protrusion, velocity, and transition time. Results indicate differences between these three measures for the production of /u/ as compared to /y/ as well as differences when upper lip movement is directed away from its target protrusion position as compared to when it is directed toward its target protrusion position. This study also includes an attempt to determine onset of protrusion in a consonant cluster followed by a rounded vowel. Results show that the protrusion movement due to the rounding of the second vowel of a group –VCC ... CV– starts as early as the first consonant of a cluster of 4–6 consonants. In some cases, it even starts during the vowel preceding the cluster. These results do not support Kozhevnikov and Chistovich’s model of a CC ... CV-type syllable. On the other hand, they are compatible with Henke’s articulatory model of speech production.

Computer-assisted measurement system for x-ray films of the vocal tract.

Measurements on sagittal X rays of the vocal tract may be made directly in terms of the variables of an articulatory model by superimposing a computer-generated vocal-tract outline onto the x-ray image. Previous methods require an operator to trace a pen or pointer coupled to a recording device along the outlines or the projected image. The new system uses two video signals, one derived from a television camera scanning the projected x-ray image, the second from data in a computer display buffer. The computer display is modified through adjustment of control potentiometers assigned to the model variables. Convergence to a match between the two images is found to be a substantially quicker and easier task than manually tracing the image outline.

Articulatory model for the study of speech production.

Study of midsagittal x-ray tracings reveals that the vocal-tract outline can be accurately represented by means of variables specifying the positions of the jaw, tongue body, tongue tip, lips, velum, and hyoid. As the articulators move, they modify the vocal-tract cross-sectional area and the vocal-tract transfer function computed thereform. The speech signal may be synthesized by concatenating the responses to repeated excitation of the quasistatic vocal tract. Vowels are specified in terms of variables denoting the positions of the jaw, tongue body, lips, and velum. Consonants are implemented as transformations on the underlying vowel-derived articulatory states that satisfy given constraints on the position of an articulator relative to the fixed structures. The set of states which satisfy the given constraint corresponds to the allowed productions of the consonant. Coarticulation effects control the selection of the underlying state and thus determine the particular consonant produced. Vowel-consonant-vowel sequences generated with the aid of rules for articulator movement and the articulator-position to vocal-tract cross-sectional-area transformation are intelligible and exhibit coarticulation in agreement with acoustic measurements.

Models of Articulatory Dynamics

We present comparisons between the x-ray data of R. A. Houde [PhD Thesis, Univ. of Michigan (1967)] and actions of a dynamic articulatory model. The data can be fitted in two ways: (1) a model with minimum number of degrees of freedom, but with time-varying “characteristic filters” (Coker and Fujimura, J. Acoust. Soc. Amer. 40, 11(A) (1966)]; or (2) one in which different speeds of an articulator are produced by additional “characteristic variables” that would be redundant in the static case. The phenomenon that the constriction tends to move forward during /g/ was suggested by Houde to result from pressure buildup in the enclosed cavity. We treat the effect as a peculiarity in “programming” of /g/. Horizontal motions toward /g/ from front vowels are commanded 50 msec early, from back vowels, 50 msec late, thus producing a circular Lissajous pattern of motion. The dilemma of cause and effect (whether continued voicing forces tongue-body movement, or whether movement is explicitly commanded, to sustain voicing) leads to interesting questions when we attempt to generalize to other phonemes. Does circular motion occur in nasals and voiceless stops as well as /g/? Does the tendency for /g/ to devoice in some languages derive from differences in tongue control? These and other topics are discussed.

Computer-Assisted Measurement System for X-ray Films of the Vocal Tract

Measurements on sagittal x rays of the vocal tract may be made directly in terms of the variables of an articulatory model by superimposing a computer-generated vocal-tract outline onto the x-ray image. Previous methods require an operator to trace a pen or pointer coupled to a recording device along the outlines of the projected image. The new system uses two video signals, one derived from a television camera scanning the projected x-ray image, the second from data in a computer display buffer. The computer display is modified through adjustment of control potentiometers assigned to the model variables. Convergence to a match between the two images is found to be substantially quicker and easier task than manually tracing the image outline.

Determination of Articulatory Parameters from the Speech Signal

A speech analysis method is presented for obtaining a description of the vocal-tract shape in terms of articulatory parameters. A slightly modified version of the Stevens and House three-parameter articulatory model is used to specify the area function of the vocal tract. The area function is described by five parameters, namely, the area at the place of maximum constriction, the area of the lip opening, the length of the lip section, and the length of the vocal tract. Two additional parameters are used to characterize the source. Considering the vocal tract as a lossy acoustic tube, the parameters of a 12th order linear predictor for the synthetic speech signal at the output of the tube are obtained. The parameters so obtained are used to predict the current value of a speech sample by forming a linear combination of the preceding 12 samples of natural speech. The mean-squared difference between these predicted values and the actual sample values in natural speech is minimized by iteratively adjusting the seven vocal-tract and source parameters. The parameter values that minimize the error are finally used to describe the vocal-tract shape.

Acoustical consequences of lip, tongue, jaw, and larynx movement.

An articulatory model is presented. It defines a procedure for deriving a set of formant frequencies from information on the state of the lip muscles, the position of the jaw, the shape and position of the tongue body, and larynx height. The acoustic and auditory consequences of varying these parameters individually are reported. The introduction of the jaw as a separate parameter—a feature not used in previous articulatory models—makes it possible to explain why “openness” occurs as a universal phonetic feature of vowel production. According to the explanation proposed, the degree of opening of a vowel corresponds to a position of the jaw that is optimized in the sense that it cooperates with the tongue in producing the desired area function. Such cooperation prevents excessive tongue shape deformation. Our results suggest that, in order to reflect this principle of articulatory synergism, “tongue height,” although primary with respect to its acoustic consequences, should be represented as a derived feature characteristic of the final vocal-tract configuration.

Automatic System for Synthesis-by-Ruale

A fully automatic system for conversion of printed text into speech sounds is described. The system is divided into two parts: one part generates prosodic information from programmed rules, and the second part synthesizes the sounds. The sound synthesizer is realized by a programmed articulatory model of the vocal tract [C. H. Coker, J. Acoust. Soc. Amer. 47, 49 (1969)]. This paper focuses mainly on details of prosodic-feature assignment. This part of the system contains a syntax analyzer and a dictionary. Using information from the dictionary, the syntax analyzer assigns a phrase category to each word (subject, verb, object, etc.) and a potential for a pause at every phrase boundary. Each word is then assigned a stress and pitch contour using the pause potentials obtained from the syntax analyzer. In the last stage of the prosody generation, each phoneme is assigned duration, pitch, and intensity controls, compatible with the articulatory model program. In the articulatory model, coarticulation which depends upon phoneme duration is automatically treated. The prosodic-rule program is implemented for time sharing on an on-premises GE-635 computer. This program is linked via a data-communications set to the synthesizer implemented on a DDP-516 laboratory computer.

Articulatory Timing in Speech Production; Articulatory Distinctive Features

Controversy over “distinctive features” of phonemes—either conceptual, perceptual, or articulatory—has continued for a number of years, mainly for lack of definitive experiments. A feature hypothesis based on articulatory constraints is testable, however, through speech synthesis. In a synthesis-by-rule system based on an articulatory model (Coker; Boston, 1967; Kyoto, 1968), several of the model coordinates found useful for simplicity of synthesis rules correspond to distinctive features in current literature. Target values for most model coordinates are, indeed, discrete valued, either binary or ternary. Additional information is required for these discrete coordinates, however. Since articulator transition times are slow compared to phoneme durations, it is necessary, for even marginally acceptable speech, to stagger transition times of coordinates according to a scale of priorities for phonemes. This important/unimportant (“marked/unmarked”) distinction seems to be more multivalued and dialect dependent than do the target values. Moreover, there are asymmetries in the priority list, with respect to the beginning and end of the phonemes. The feature set implied by the present synthesizer rules is discussed. The results of this set of rules and the consequences of alternatives are demonstrated.