A training method for hidden Markov model with maximum model distance and genetic algorithm

Abstract

Maximum model distance (MMD) is a discriminative algorithm developed for training the whole HMM models. It differs from the traditional maximum-likelihood (ML) approach through comparing the likelihood against those similar utterances and maximizes their likelihood differences. Combined with MMD, this paper proposes a hybrid training method based on the genetic algorithm (GA). Experimental results from the TI46-Word alphabet database show that this algorithm has better performance than MMD. The reason is that the MMD algorithm is exploring only one local maximum in practice while the GA operations in the hybrid algorithm provide the ability to explore several local maximums and hopefully the global maximum.