New method of speech signals adaptive features construction based on the wavelet-like transform and support vector machines

Abstract

In this paper it is proposed the original method of adaptive feature vector construction for speech signals based on wavelet transform and support vector machines (SVM). For wavelet basic function generation it was proposed to use genetic algorithm and SVM based classification accuracy as the objective function. It was shown that the usage of the generated in such a way wavelet functions lets to improve speech signals classification accuracy. In particular the accuracy improvement is 1% to 5% in comparison with mel-frequency cepstral coefficients.