Recognition of demisyllable based units using semicontinuous hidden Markov models

The authors describe the use of semicontinuous hidden Markov models (SCHMM) for recognition of demisyllable based units within a speaker-dependent automatic recognition system for continuous speech. The processing units are demisyllables, while the decision units consist of consonant clusters and vowels. During recognition, the Viterbi algorithm is used for implicitly localizing the syllable boundaries. The estimation of the model parameters is achieved by the Viterbi training algorithm combined with a simple procedure for generating seed models. The basic principles of the algorithms are presented in detail. Application of the SCHMM approach resulted in a significantly higher performance than using discrete HMMs. The experimentally evaluated recognition rates are discussed with respect to some simplifications in the training and recognition algorithms