The problem of the optimal speech feature selection with respect to CDHMM discrete-utterance recognition is addressed in the paper. Two sequential search algorithms - the Sequential Forward Search (SFS) and Sequential Floating Forward Search (SFFS) have been investigated and applied to several speech databases. It was observed that the dynamic (delta) parameters derived from frame energy and cepstrum were identified as more important than the static ones. The hypothesis about the high discriminative power of the dynamic features has been confirmed in a series of speaker-independent tests. We have even demonstrated that the recognition rate achieved with only the dynamic features closely approached that of the complete feature set. As a practical application of the study, we have proposed a two-level HMM classification scheme that may significantly reduce recognition time without a loss of accuracy.
Bibliographic reference. Nouza, Jan (1995): "On the speech feature selection problem: are dynamic features more important than the static ones?", In EUROSPEECH-1995, 919-922.