Data-driven temporal filters based on maximum mutual information for robust features in speech recognition

Abstract

Linear discriminant analysis (LDA), principal component analysis (PCA) and minimum classification error (MCE) have been used to derive data-driven temporal filters in order to improve the robustness of speech features for speech recognition. In this paper, the criterion of maximum mutual information (MMI) is proposed for constructing the temporal filters, and detailed comparative analysis among these various approaches are presented and discussed. Experimental results show that the MMI-derived temporal filters significantly improve the recognition performance of the original mel frequency cepstrum coefficients (MFCC) features compared to LDA/PCA/MCE-derived filters. Also, while the MMI-derived filters are combined with the conventional temporal filters, cepstral mean and variance normalization (CMVN), the recognition performance can be further improved.