An Environment Model-based Robust Speech Recognition

摘要

In this paper, a new approach named environmental discrimination learning (DEL) is proposed to remove the effects of the environment noises including the additive noise and the channel distortions. This method optimizes the environment parameters by the minimum classification error (MCE) criterion which trains the parameters of a given class dependently on the whole classes. The EDL approach utilizes more about hte information between different classes to optimize the environment parameters, therefore, it can minimize the error rate. And a generalized probabilistic descent (GPD) algorithm is adopted for discriminative training the environment parameters. A speaker independent isolated word recognition system based on whole word-HMM model is used to evalaute the proposed approach. Experimental resutls show that the proposed method achieves significant improvement of recognition performance.