A New Method for MLE Training Based on Multi-model Weighting

A new method based on multi-model weighting for maximum likelihood estimation (MLE) is proposed in this paper. In order to ease the assumptions of maximum likelihood training, the model is generated based on the weight of multi-model which were trained with the divided training data respectively, the weight is gained according to the principle that the higher ratio of inter-variance to intra-variance of the class, the better discriminative the model is, therefore a greater weight would give to it, then the new models will be more discriminative than the original models. The experiments on speech recognition demonstrate that the new model out-performed the model that trained with traditional method.