Modified Linear Discriminant Analysis for Speech Recognition

In this paper, a new method for extracting discriminant features in automatic speech recognition (ASR), termed modified linear discriminant analysis (MLDA), is proposed. As a generalization of linear discriminant analysis (LDA), MLDA integrates the cluster information in each class by redefining the between-class scatter matrix based on the fact that many clusters exist in each state in hidden Markov model (HMM)-based ASR. Experimental results on TiDigits show that our presented MLDA clearly outperforms LDA and clustering-based linear discriminant analysis (CLDA), which was proposed for facial expression recognition, and about a 10% string error rate reduction (SERR) is found.