A model block-training method for HMM-based speech recognition systems

A model block-training method for hidden Markov models (HMMs) is described. It combines several model estimations from several training sets, all of which are derived from utterances of the same word, into a new one. Although the recognition rate of the recognizer trained by the block-training method is lower than that of a recognizer trained by a batch-training method, the new one is much more flexible than the old one in building up the multispeaker or speaker-independent system in spite of the significant difference in spectral features. A computer simulation shows that the block-training procedure can converge to a local maximum point of likelihood function.<
>