Probabilistic vector model for voicing mode identification of intervocalic stop consonants

Abstract

A probabilistic vector model was developed to identify the voicing mode of phonemically equivalent intervocalic stop consonants with an accuracy commensurate with trained listeners, such performance to be attained independently of the succeeding vowel's identity or the differences among talkers. Acoustic features known to be salient to the perception and production of stop consonants were studied as random processes whose probability density functions provided models for voicing mode identification. Each acoustic feature present in a stop's production then contributed a vector whose magnitude and direction were determined from the probability distributions resulting from these density functions. Stop recognition was the result of summing these individual vectors, with the resultant vector specifying the voicing mode of the stop under consideration. Acoustic features incorporated into the model were closure duration, stop-consonant duration, fundamental frequency, adjacent vowel amplitudes, burst amplitude, formant transitions, voicing during closure, and voice onset time. When the performance of the probabilistic vector model was compared with that of trained listeners, correct voicing mode identification was approximately 99.0% in both cases.