Hidden Markov modelling of speech based on a semicontinuous model

The discrete hidden Markov model (HMM) is extended here by using a combination of continuous Gaussian density functions derived from a vector quantised codebook together with discrete HMM output probabilities. Experimental results for a vocabulary consisting of the digit set for a group of speakers have shown that the semicontinuous approach to HMM offers improved performance in comparison to both the discrete HMM and to dynamic time warping methods which incorporate template adaptation