Wavelet transform based automatic speaker recognition

Abstract

An effective feature extraction technique for speaker recognition is presented in this paper. It uses multiresolution property of wavelet transform and Mel-Frequency Cepstral Coefficients (MFCCs) for analyzing the speech signal. For individual speaker, first the speech signal is decomposed using Discrete Wavelet Transform (DWT) into approximations and details coefficients. Approximation coefficients are then used to compute MFCCs. Experimental results were computed on PIEAS Speech Database for text independent speaker identification. The proposed method gives very good recognition rate i.e. 96.25% for non telephonic and 86.77% for telephonic speech data. In addition to this, analysis for choosing the appropriate number of MFCCs, the appropriate number of decomposition levels and wavelet type has also been performed.