Speaker recognition based on wavelet packet decomposition and Volterra adaptive model

Abstract

The process of voice generation belongs to nonlinear system, and the voice signal is chaotic. The traditional Volterra model is generally 2 order truncation, a low order filter is used to estimate the speech signal, and prediction effect is not accurate. So, this paper proposes a feature extraction scheme based on the wavelet packet decomposition and Volterra adaptive model. Firstly, the speech signal will be decomposed by wavelet packet. Secondly, reconstruct the phase space for all sub-band signals. Thirdly, using second order Volterra series expansion and the linear adaptive FIR filter to estimate parameter vector H(n) and output signal U(n) for Volterra model, and weight vectors of Volterra filter are obtained for speaker recognition. Speaker recognition experiment is completed based on hidden Markov model. The experimental results show the extracted features have been obviously improved, especially in the noise environment.