Effects of token variability on resolution for vowel sounds.
Rosalie M. Uchanski, Kathleen M. Millier, D. E. Ronan, C. M. Reed, Louis D. Braida · The Journal of the Acoustical Society of America · 1991
Variability in the production of speech elements complicates the recognition problem for both humans and machines. Paradoxically, variability from token to token of a given speech sound also makes it possible to discriminate between fixed tokens of the sound. Although some components of this variability have no linguistic significance, they may serve as cues when identifying or discriminating specific vowel tokens. The results are reported of identification and discrimination experiments in which the number of tokens used to represent each speech sound was varied. The speech sounds consisted of ten monophthongonal vowels spoken by a single speaker in /h/–V–/d/ context. Generally measures of resolution for vowel contrasts decrease as the number of tokens increases, but the reduction is greater when the number of tokens is increased from one to four than from four to sixteen. The analysis suggests that when one or four tokens of each vowel are used in discrimination tasks, listener judgments are based on perceptual qualities specific to the tokens rather than to general characteristics of the vowels. [Work supported by NIH.]