Efficient two-stage vector quantization speech coder using wavelet coefficients of excitation signals

Abstract

An improved backward prediction coder featuring two-stage vector quantization (VQ) of shape codevectors is presented. Efficient two-stage VQ is achieved using the wavelet coefficients of excitation signals; i.e., wavelet coefficients are calculated by applying a discrete wavelet transform to excitation signals, and the results are divided into an approximation group and a detail group. The data lengths of both approximation and detail coefficients are half that of conventional two-stage VQ systems. Simulation results show that the proposed coder achieves a better weighted signal-to-noise ratio (WSNR) than conventional coders and, in terms of reconstructed speech quality, ranks between the FS-1016 Code Excited Linear Prediction (CELP) coder and the Vector Sum Excited Linear Predictive Coding (VSELP) coder.