A robust formant-based speech spectrum comparison measure

Accurately labeled (as F1, F2etc.) sets of formant frequencies are known to be good indicators of the phonetic identity of a speech sound and thus form an attractive acoustic representation for speech recognition. Labeling errors are, however, common, and they make the pairwise comparison of sets of formant frequencies an unreliable process. In this paper a dynamic-programming-based sequence comparison algorithm is proposed which allows unlabeled sets of formant frequencies to be compared. Preliminary speech recognition experiments using this algorithm give results comparable with those obtained on the same data using conventional spectrum matching. In view of the likely potential for improvement in the formant-based method, the results are considered encouraging. Applications to other problems involving speech analysis are also discussed.