Discrete Wavelet Transform for automatic speaker recognition

Abstract

This paper deals with automatic speaker recognition. We consider here a context independent speaker recognition task with a closed set of speakers. We have shown in [1] a comparative study about the most frequently used parametrization/classification methods for the Czech language. Wavelet Transform (WT) is a modern parametrization method successfully used for some signal processing tasks. WT often outperforms parametrizations based on Fourier Transform, due to its capability to represent the signal precisely, in both frequency and time domains. The main goal of this paper is thus to use and evaluate several Wavelet Transforms instead of the conventional parametrizations that were used previously as a parametrization method of automatic speaker recognition. All experiments are performed on two Czech speaker corpora that contain speech of ten and fifty Czech native speakers, respectively. Three discrete wavelet families with different number of coefficients have been used and evaluated: Daubechies, Symlets and Coiflets with two classifiers: Gaussian Mixture Model (GMM) and Multi-Layer Perceptron (MLP). We show that recognition accuracy of wavelet parametrizations is very good and sometimes outperform the best parametrizations that were presented in our previous work.