Given that traditional method of hidden Markov models parameter initialization for speech recognition (random method,kmeans) can lead to convergence in local optimization of model parameters rather than global optimization problems.A new approach was proposed with three steps.First,the initial center was selected according to the maximum distance; second,the original data was split into small kinds by the minimum distance;finally,the interference point in the small kind was eliminated.The method resulted in much more similarity kmeans than traditional method in the kind.Experimental results show that the improved method has the better approximation of smooth voice characteristics and improves the speech recognition rates which comparing with traditional methods.