Combined scalar-vector quantization: a new spectral coding method for low rate speech coding

Abstract

Linear prediction is the dominant model in low rate speech coding and line spectral frequencies (LSFs) are often used as parameters to represent the vocal tract filter in speech coders using linear prediction. This paper proposes a new vector quantization method for quantization of the LSFs: namely combined scalar-vector quantization (CSVQ). It is shown that this spectral coding method requires negligible computation overhead compared to scalar quantization, which is far less than other VQ schemes, even the fast quantization techniques, such as tree-searched vector quantization. Several codebook training algorithms are suggested in this article. Results of experimental simulations verify the satisfactory performance of the new proposed vector quantization method.