Tree-structured vector quantization of speech lsf parameters
D. Chemla, Sau-Wah Soong, Wai-Yip Chan · 2005
Multistage tree-structured vector quantization (MSTVQ) of speech linear prediction filter parameters is evaluated, aiming to obtain good distortion-rate performance at low en coding search complexity. For each speech analysis frame, the coefficients of the tenth order linear-prediction filter are represented as line spectral frequencies and intraframe coded using several stages of binary tree-structured VQ (TSVQ) codebooks. Experimental performance data are gathered for various degrees of codebook storage and encoding search complexity by varying respectively the number of stages and the number of encoding-search survivors. The results show that MSTVQ can furnish transparent coding quality for rates between 23-25 bits per frame. The required encoding complexity ranges from below 200 to several hundred weighted squared error distortion computations, and the required stor age complexity ranges from below 200 to about 1500 code vectors.