Sub-band Information Fusion Based on Wavelet Thresholding for Robust Speech Recognition

In recent years, sub-band speech recognition has been found useful in addressing the need for robustness in speech recognition, especially for the speech contaminated by band-limited noise. In sub-band speech recognition, the full band speech is divided into several frequency sub-bands, with the result of the recognition task given by the combination of the sub-band feature vectors or their likelihoods as generated by the corresponding sub-band recognizers. In this paper, we draw on the notion of discrete wavelet transform to divide the speech signal into sub-bands. We also make use of the robust features in sub-bands in order to obtain a higher sub-band speech recognition rate. In addition, we propose a likelihood weighting and fusion method based on the wavelet thresholding technique. The experimental results indicate that the proposed weighting methods for likelihood combination and classifiers fusion improve the sub-band speech recognition rate in noisy conditions.