Two-Stage Vector Quantization Based Multi-band Models for Speaker Identification

This paper presents an effective method for speaker identification. Based on the wavelet transform, the input speech signal is decomposed into several frequency bands, and then the linear predictive cepstral coefficients (LPCC) of each band are calculated. Furthermore, the cepstral mean normalization technique is applied to all computed features in order to provide similar parameter statistics in all acoustic environments. We propose a multi-band 2-stage vector quantization (VQ) as the recognition model in which different 2-stage VQ classifiers are applied independently to each band and the errors of all 2-stage VQ classifiers are combined to yield total error and a global recognition decision. The experimental results show that the proposed method gives better performance than other recognition models proposed previously in both clean and noisy environments.