Sub-band weighted projection measure for robust sub-band speech recognition

Abstract

In recent years, sub-band speech recognition has been found useful in robust speech recognition, especially for speech signals contaminated by band-limited noise. In sub-band speech recognition, full band speech is divided into several frequency sub-bands and then sub-band feature vectors or their generated likelihoods by corresponding sub-band recognizers are combined to give the result of recognition task. In this paper, we concatenate sub-band feature vectors, where we extract phase autocorrelation (PAC) MFCC, as noise robust features, from each sub-band. Furthermore, we extend a model adaptation method, named sub-band weighted projection measure (SWPM), to adapt HMM Gaussian mean vectors to concatenated sub-band feature vectors in noisy conditions. The experimental results indicate that the proposed method significantly improves the sub-band speech recognition system performance in presence of additive noise.