A noise-robust speech recognition system based on wavelet neural network

Abstract

Aiming at the problem that the performance of speech recognition system will drop severely in noisy environments, this paper proposed a recognition system that has excellent anti-noise performance. The feature parameters of front-end are ZCPA (Zero-Crossings with Peak-Amplitudes) feature and the recognition network of back-end is wavelet neural network. Mexican Hat wavelet was used to replace the Sigmoid or Gaussian basis function of feed-forward neural network. The network structure was three layers. Training network weights, determining the position and scale factor of wavelet function were processed respectively. And the information of training samples was made full use of when estimating the position factor. Hence, the network could converge rapidly and it also avoided the problem of wavelet "dimension disaster". A 50 isolated-word person-independent speech recognition system using BP network or wavelet neural network as recognition network was simulated under different SNRs in this experiment. The experimental results showed that recognition rates using wavelet neural network are higher than using BP network. © 2011 Springer-Verlag.