Evaluation of a speaker-dependent recognition metric as a substitute for human judgments of speech quality
C. S. Watson, Diane Kewley-Port, Daniel P. Mäki, D. Reed · The Journal of the Acoustical Society of America · 1987
The Indiana Speech Training Aid project (ISTRA) is evaluating the use of speaker-dependent speech recognition to provide speech-drill feed-back for deaf or misarticulating children. The ISTRA systems employ IBM PC's with Interstate Voice Products Vocalink SRB boards. The basic approach is to form templates from the child's best current productions. Feedback in drill sessions is then based on a goodness metric. That metric represents the “match” between 10-ms sampling of a new utterance and the corresponding information for the stored template. The goodness metric has been partially validated as a substitute for feedback judgments provided by a human teacher. In general, the method compares values of the goodness metric to actual ratings by a jury of listeners. For words spoken by either normal speakers who intentionally vary the quality of their pronounciations, or by hearing-impaired speakers, average inter-listener correlations for ratings of quality did not differ from the correlations between the goodness metric and the average listener's judgements. Test-retest reliability is, of course, much higher for computer than for the human listener. [Work supported by NSF.]