Robust features for speech recognition using minimum variance distortionless response (MVDR) spectrum estimation and feature normalization techniques

摘要

In this paper, feature extraction methods based on frequency-warped minimum variance distortionless response (MVDR) spectrum estimation are analyzed and tested. The effectiveness of the conventional FFT-based mel-frequency cepstrum coefficients (MFCC) and the MVDR-based features are carefully compared. Two normalization techniques are further applied to improve the robustness of the features: the widely used cepstral normalization (CN), and newly proposed progressive histogram equalization (PHEQ). Extensive experiments with respect to the AURORA2 database were performed. The results indicated that both the MVDR-based features and the normalization processes are very helpful.