Multiband pitch and residual coding of speech signals
Paul G. Mermelstein, P. Zeng, M. Saikaly, Yasheng Qian · 2002
We explore improvements to code-excited linear prediction of speech coders near 7 kb/s that allow attainment of toll quality while maintaining low complexity. We introduce three-band pitch predictors to increase pitch prediction gain through the use of three pitch gain parameters, one for each frequency band. Prediction of the gains in the higher bands from that determined for the lowest band avoids the need to increase the transmitted information. The same three-band structure is exploited for codebook index generation by segmenting a full-band trained codebook into three band-pass components. Very small size codebooks suffice for the bands above 1 kHz, allowing relatively more bits to be assigned to the lowest codebook. Subjective evaluation results indicate that in channels without errors the proposed codec achieves a speech quality comparable to the 32 kb/s ADPCM standard G.726, i.e., toll quality.