Use of Multifrequency Channel Decomposition for Speech Recognition

In speech recognition, recognition performance is usually affected by the confusing set in the vocabulary. One possible way to improve the discriminability is to decompose speech signal into multifrequency channels with different weights. In this paper, speech signal is decomposed into multiple spatial frequency channels using wavelet transform and filter banks, respectively. Speech signal in each channel is then used to calculate the LPC-derived cepstral coefficients. For each channel, a Bayesian network is adopted to model speech features. Finally, a channel weighting method is used to emphasize the contributions of different channels. For experimental evaluation, recognition of English E-set and 200 city names provided by 20 speakers were used to evaluate the proposed method. The experimental results show that multifrequency channel decomposition approach achieves a better performance compared to the conventional single-channel method. In addition, the wavelet transform and filter bank approaches have comparable performance.