Semi-continuous hidden Markov models in isolated word recognition

Abstract

A semicontinuous hidden Markov model is proposed to incorporate the vector quantization distortion into the general hidden Markov model methodology under a probabilistic framework. It provides a relatively simple but powerful tool for modeling time-varying signal sources. Experimental results show that the recognition accuracy of the semi-continuous model is measurably improved in comparison to that of the conventional discrete hidden Markov model and template-based dynamic time warping techniques.<>