An approach towards a synthesis-based speech recognition system

Preliminary results are presented of experiments with a recognition scheme intended for continuous speech. The scheme utilizes information about interphoneme contextual effects contained in formant transitions and employs internal trial synthesis and feedback comparison as a means for recognition. The aim is to achieve minimal sensitivity to the appreciable variability which occurs in the speech signal, even for utterances of a single speaker. While the approach outlined here is quite general, it has initially been tried out on vowel-stop-vowel utterances. Recognition scores obtained are encouraging and demonstrate the viability of the approach.