A New Method for Discriminative Model Combination in Speech Recognition

A new method based on discriminative model combination for acoustic model training is proposed. The MPE trained model and the MMIE trained model are used for model combination. The combination criterion is based on the ratio of the inter-variance to the intra-variance of each model. Besides we also propose a ldquoclusterrdquo method for the model to choose its confusion models in order to get the accuracy information for the combination weight calculation. The experiments demonstrate that the new model can get better performance than any of the single MPE and MMIE trained models.