Abstract: Traditional training methods such as maximum likelihood estimation (MLE) do not consider discriminative relation between acoustic models, so some models are apt to obscure each other. In order to raise the differentiation degree between models, discriminative training criteria are proposed. This paper mainly introduces the new discriminative training method, the minimum phone error (MPE), and proposes a new model combination method. For the single mixture model, it weights the models of MPE and MLE based on the confusion of them. For the multimixture model, a model selection method is proposed. Experiments demonstrate that this method achieves consistent improvement over the conventional MPE training and the relative error reduction is about 4% for the single mixture model and about 3% for the multi mixture model.