Predictive SVD-transform coding of speech with adaptive vector quantization

Abstract

A predictive transform coder using singular value decomposition (SVD) and adaptive vector quantization (AVQ) is presented. The linear predictive coding (LPC) excitation is obtained by a weighting VQ quantization of the SVD transforms of the prediction residual. The orthogonal representation of the excitation provided by the SVD transform permits a fine quantization of the perceptually important components. Some dynamic bit allocation approaches based on the relative importance of the singular values are described. The integration of the usual perceptual noise-shaping is easy, and a simple noise-shaping based on the direct reduction of the singular values spread is presented. The LPC coefficients are quantized by an adaptive multistage VQ approach. The results show good performance within the range of 7-8 kb/s