Hidden Markov Models are very successful in modeling the acoustic behavior of speech.They may be classified into two groups

continuous HMMs (CHMMs) and discrete HMMs (DHMMs)

according to the output probability distribution.Traditional training methods such as Baum-Welch algorithm are noted for the rapid convergence.However

these methods are hill-climbing based algorithms and they just lead to locally optimal solutions

which might deteriorate the recognition rate.For CHMMs

a segmental k-means method has been developed to get reliable initial estimate and thus guaranteed the rapid and proper convergence.For DHMMs

this offers little help and the final solution is a locally optimal solution.While one outstanding character of evolutionary computation is global search

it can converge to a globally optimal solution or at least a sub-optimal solution.In this paper evolutionary computation is applied to training DHMMs.A hybrid training method that combines the traditional method and evolutionary computation is proposed.Experimental results show that the proposed method has both qualities of global search and rapid convergence and the resulting models are superior to those obtained with traditional methods or simple evolutionary computation and eventually contribute to the increase of recognition rate.