Classes Of Speech Sounds Research Articles

Analysis of Acoustic Speech Signals into Articulatory Parameters

An analysis technique is discussed whereby speech spectra are compared with spectra generated by a computer-simulated model of articulation. In this model, a spectrum is determined by specifying as input parameters a vocal-tract area function (or a simplified configurational description) and the source spectral envelope and location. The input parameters are adjusted to obtain a “spectral match” between the internally generated spectrum and the speech spectrum, thereby specifying the articulatory parameters that correspond to the given speech spectrum. Data relating to matches from several classes of speech sounds are presented. Questions pertaining to the uniqueness and applicability of the parameters obtained by this procedure are discussed, with particular reference to descriptions of vocal-tract dynamics. The possibility of using parameters derived by this technique to control a dynamic vocal-tract analog synthesizer is also considered. [This work was supported in part by the U. S. Army, the U. S. Air Force Office of Scientific Research, and the U. S. Office of Naval Research; in part by the U. S. Air Force (Electronic Systems Division) under Contract AF 19(604)-6102; in part by the National Science Foundation (Grant G-16526); and in part by the National Institutes of Health (Grant MH-04737-02).

Speech Sound Classification Based on Signal Statistics

A method of quantizing the output of a vocoder-type speech analyzer is described. This quantization effectively classifies the sound at each instant into one of a small number (typically, less than 100) of classes, but in such a way as to allow relatively accurate reconstruction of the original unquantized voltages The boundaries of the classes are chosen on the basis of the statistical distribution of the voltages, using redundancies which may exist to reduce the number of classes to a low value. Experimental results are presented, illustrating classification schemes resulting from this procedure. These schemes bear some resemblance to conventional phonetic analyses of speech sounds, but they have the advantage of being derivable by purely mechanical means, without the use of human intuition or judgment. The use of such sound classification for automatic speech recognition is discussed.

Noise Duration as a Cue for Distinguishing among Fricative, Affricate, and Stop Consonants

An earlier experiment [Gerstman, Liberman, Delattre, and Cooper, J. Acoust. Soc. Am. 26, 952 (1954)] demonstrated that variations in the duration of a change in formant frequency can serve as an effective cue for distinguishing one class of speech sounds from another. The duration cue was found to operate independently of place cues; i.e., it served to distinguish the stop b from the semivowel w, both of which are articulated at a speaker's lips. Exploratory work now suggests that another temporal cue, noise duration, operates in much the same fashion, distinguishing among classes of sounds produced at the same place in the mouth. A tape recording of the spoken syllable ∫a (as in shop) was progressively shortened by cutting off increasing amounts of the noise portion with which the syllable begins. When the noise portion was approximately half its original length the syllable was heard as ∫a (as in chop), and when most of the noise had been removed the syllable was heard by some listeners as ka and by others as ta. (Inasmuch as ∫a and t∫a are both produced at a point in the mouth that lies be|tween the places where k and t are formed, the listeners' uncertainty as to whether the stop is k or t is understandable.) Thus the consonant, originally a fricative, became an affricate and then a stop. The experiment has been repeated with more highly controlled synthetic speech patterns. Here it was found that a second cue of some importance was the time required by the friction to rise to full intensity. [This work was supported in part by the Carnegie Corporation of New York, and in part by the Department of Defense in connection with Contract DA49-170-sc-1642.]

Rate and Duration of Change in Formant Frequency as Cues for the Identification of Speech Sounds

Earlier experiments have shown that the direction and extent of the frequency changes in the formants of synthetic speech sounds can serve as effective cues for the identification of certain phones according to place of production. Examination of spectrograms and exploratory work with synthetic speech has suggested that the temporal course of these changes may provide a basis for distinguishing among some classes of speech sounds that differ in manner of production. It was found in the present experiment that the tempo of the change between two fixed formant frequencies is, in fact, sufficient to enable naive listeners to distinguish rather reliably among stop-consonant-plus-vowel (e.g., ba), semi-vowel-plus-vowel (e.g., wa), and a sound of shifting vowel color (e.g., ua). Further experiments suggest that the duration of the frequency changes are more important than the rate at which the changes occur.