In speech recognition systems based on Hidden Markov Modeling, the computation of the likelihoods in detailed models is intensive, while the performance of crude models is poor. A hybrid model which combines the detailed and crude models is proposed to take advantage of the performance of the detailed model and the speed of the crude model. Experimental results show that a significant (up to a factor of 20) likelihood computation reduction has been obtained, with almost the same recognition accuracy as the baseline models on both the speaker-dependent and speaker-independent systems.
Bibliographic reference. Li, Z. / Kenny, P. / O'Shaughnessy, Douglas (1995): "Hybrid hidden Markov models in speech recognition", In EUROSPEECH-1995, 795-798.