Abstract: This paper proposes a hybrid MMI/EC architecture to be embedded in the training of HMMs. Each individual in evolutionary computation represents a set of HMMs, while the fitness value of each individual represents the maximum mutual information. Since evolutionary computation is noted for its global search and population-based operations, the globally optimal solutions (or the sub-optimal solutions) can be obtained. Experiments indicate that the system trained with the proposed method is superior to the one trained with the traditional gradient based training method.

Key words: speech recognition, evolutionary computation, maximum mutual information