Wavelet packet cepstral analysis for speaker recognition

Abstract

A novel processing technique for speaker recognition applications is introduced. It was shown that feature extraction based on cepstral analysis of the wavelet packet decomposition can provide significant inter-speaker separation. This idea is based on deconvolution of the vocal tract and excitation source components through homomorphic decomposition of a signal's multiresolution wavelets. A simple neural network technique is employed to classify the feature vector obtained through wavelet packet cepstral analysis.