New methods for robust speech recognition

Abstract

New methods of feature extraction, end-point detection and speech enhcincement are developed for a robust speech recognition system. The methods of feature extraction and end-point detection are based on wavelet analysis or subband analysis of the speech signal. Two new sets of speech feature parameters, SUBLSF’s and SUBCEP’s, are introduced. Both parameter sets are based on subband analysis. The SUBLSF feature parameters are obtained via linear predictive analysis on subbands. These speech feature parameters can produce better results than the full-band parameters when the noise is colored. The SUBCEP parameters are based on wavelet analysis or equivalently the multirate subband analysis of the speech signal. The SUBCEP parameters also provide robust recognition performance by appropriately deemphasizing the frequency bands corrupted by noise. It is experimentally observed that the subband analysis based feature parameters are more robust than the commonly used full-band analysis based parameters in the presence of car noise. The a-stable random processes can be used to model the impulsive nature of the public network telecommunication noise. Adaptive filtering are developed for Q-stable random processes. Adaptive noise cancelation techniques are used to reduce the mismacth between training and testing conditions of the recognition system over telephone lines. Another important problem in isolated speech recognition is to determine the boundaries of the speech utterances or words. Precise boundary detection of utterances improves the performance of speech recognition systems. A new distance measure based on the subband energy levels is introduced for endpoint detection.