Robust Speech Recognition Method Based on Discriminative Evironment Feature Extraction

It is an effective approach to learn the influenceof environmental parameters, such as additive noise and channeldistortions, from training data for robust speech recognition. Most ofthe previous methods are based on maximum likelihoodestimation criterion. However, these methods do not lead to a minimumerror rate result. In this paper, a novel discriminative learningmethod of environmental parameters, which is based on MinimumClassification Error (MCE) criterion, is proposed. In the method, asimple classifier and the Generalized Probabilistic Descent (GPD)algorithm are adopted to iteratively learn the environmentalparameters. Consequently, the clean speech features are estimated fromthe noisy speech features with the estimated environmental parameters,and then the estimations of clean speech features are utilized in theback-end HMM classifier. Experiments show that the best error ratereduction of 32.1% is obtained, tested on a task of 18 isolatedconfusion Korean words, relative to a conventional HMM system.