On the generation and use of a parallel-branch subunit model in continuous HMM

In this paper, we propose two methods of obtaining a parallel-branch subunit model for improved recognition performance. In one method, the model is obtained by adding a new subunit branch based on misrecognized data in training to the previous parallel branches for that submit. In the other method, it is obtained by splitting off each subunit branch based on mixture components in continuous hidden Markov model. We propose to use a good initialization point obtained by error corrective estimation rather than by random or probabilistic statistics in the parallel-branch model when the number of mixtures and parallel branches increases.