Development of a VQ-HMM continuous speech speaker-independent recognition system for small vocabularies

The authors describe the first phase in the development of a speech recognition system for small vocabularies. The system is designed to handle speaker-independent continuous speech and can be easily modified for different vocabularies. The system consists of a spectral analysis stage followed by vector quantization (VQ) and hidden Markov modeling (HMM). VQ is performed by multiple acoustic parameter codebooks, which are independent or dependent on speech units. The approach seeks to incorporate more knowledge about phoneme allophonic, linguistic, and speaker-dependent variations. The increase in number of codebooks is compensated by their decrease in size, minimizing effects on storage requirements. Phoneme HMM models permit an easy adaptation to new vocabulary requirements. A loop HMM structure with optional silences between words allows continuous speech recognition using a Viterbi search. Preliminary experiments on continuous phoneme and digit recognition were performed on an unrestricted-speaker telephone database