Split Vector Quantization of Psychoacoustical Modified LSF Coefficients in Speech Coder Based on Pitch-Tracking Periodic-Aperiodic Decomposition

Abstract

This paper presents methods of detection and quantization coefficients from speech coder based on Pitch-Tracking Periodic-Aperiodic Decomposition. Spectral envelopes of harmonic and noise components of speech are represented by linear spectral frequencies (LSF) coefficients. In this article we show new methods of quality improvement in coding signal, using perceptual properties of human ear. Spectrum envelopes of the harmonic frequencies and noise components envelopes are represented in psychoacousticaly based phon-Bark and sone-Bark scale. Line Spectral Frequencies coefficients are quantized, using methods proposed by R.M. Gray, Y. Linde and A. Buzo. For better performance, reduction of computational complexity and memory space, Split Vector Quantization methods are used. Structure of Vector Codebook is condtructed based on dependency between LSF coefficents in vectors of training sequentions. LSF vectors are patritioned into two or three sub-vectors, and each of them is coded separately. Bit realocation between sub-codebooks allow obtain improvement of codebook quality. Combination of perceptual properties and modified split vector quantization of speech signal parameters, permit achieve good quality signal in speech coder for transmission rate 2.8 kbit/s and below.

This work is supported by Bialystok Technical University under the grant WAVI/2/04