Classified vector quantization of LPC parameters

Abstract

To achieve high coding efficiency, modern speech coders adopt hybrid coding approaches, which utilize different coding mechanisms for various classified speech segments. With known voiced/unvoiced detection, in this paper, a classified LPC quantization (CLPQ) scheme is presented to effectively encode line spectral frequencies (LSF). The proposed CLPQ scheme improves the performance of the classified LSF vector quantizer, which adopts two LSF codebooks derived separately from voiced and unvoiced speech frames. With an objective spectral distortion measure, the CLPQ scheme successfully reduces the bit rate by about 1 bit/frame. Many classified LSF quantizers with different codebook structures and bit rates were evaluated. It would be helpful to design a classified LSF quantizer, which arrives at a compromise between distortion, bit rate and computational complexity.