The Research of Noise-Robust Speech Recognition Based on Frequency Warping Wavelet

Abstract

The main task of speech recognition is to enable computer to understand human languages (Lawrence, 1999; Jingwei et al., 2006). This makes it possible that machine can communicate with human. Usually, speech recognition includes three parts: pre-processing, feature extraction and training (recognition) network. In this paper, the speech recognition system is described as Fig. 1. It consists of filter bank, feature extraction and training (recognition) network. The function of filter bank is dividing speech signal into different frequency band to be good for extraction feature. The good feature can improve the system recognition rate. The training (recognition) network trains (recognizes) the feature vectors according to feature mode and outputs recognition results. The research on noise-robust capability of speech recognition system is a difficult problem that has been limiting the practical application of the speech recognition system (Tianbing et al., 2001). Because human ear has strong noise-robust capability, it is very important to abstract the features of fitting auditory characters of human ear for improving system noiserobust performance. The warping wavelet overcomes the disadvantage that the common wavelet divides frequency band in octave band and it is more suitable to the auditory characters of human ear. Bark wavelet is a warping wavelet that divides frequency band according to critical band (Qiang et al., 2000). At the same time, MFCC (Mel Frequency Cepstrum Coefficients) (Lawrence, 1999) and ZCPA (Zero-Crossing with Peak Amplitude) (Doh-suk et al., 1999) features themselves have noise-robust performance. HMM is classical recognition network, and wavelet neural network is also popular recognition network (Tianbing et al., 2001). So considering above three parts of speech recognition system, the paper used the two kinds of filters: FIR filter and Bark wavelet filter; two kinds of features: