On the Use of Different Numbers of Mixtures in Continuous Density Hidden Markov Models

Abstract

In the continuous density hidden Markov model for speech recognition, the number of mixture components in each state is usually fixed throughout all the states. The authors propose the use of a different number of mixture components for each state. For this purpose, a method is also proposed for determining the number of mixture components from the entropy information of each state. The recognition accuracy with the proposed algorithm improves considerably.