A cross-language performance study of vector quantisation

Abstract

This paper investigates the performance of split vector quantisation (VQ) of the line spectral frequencies (LSFs) across a set of 10 modern languages. Spectral quantisation accounts for a significant portion of the bit allocation in low-rate speech coding. Split VQ of the LSFs can achieve transparent quantisation of the linear prediction coefficients at 24 bits/frame. The codebooks are trained on individual languages and the cross-language VQ performance was measured using spectral distortion (SD). The results show that the spectral structure of the codebook training language influences the performance of the VQ. The number of bits/frame required for transparent speech varied by as much a 2 bits across languages.