A noise-robust speech recognition system based on ZCPA features and support vector machine

In the noisy environment, the performance of speech recognition system may become worse to some extent. In order to solve this problem, this paper used the zero-crossings with peak amplitudes (ZCPA) features as speech feature parameters, which are based on human hearings property. The extraction method of ZCPA features is that calculating the unward zero-crossing rate of speech signal gets frequency information, compressing amplitudes non-linearly incorporates intensity information, and form finally output features. This paper also used the support vector machine as recognition machine, realized a speech recognition system of non-specific person and isolated words with Visual C++ programming, and got the recognition results in different SNRs and in different words. Experiments indicate that the recognition correct rates based on ZCPA features and support vector machine are much higher than those of based on traditional hidden Markov model, and also have better robustness.