Techniques for low bit rate speech coding using long analysis frames

A time-variable LPC (linear predictive coding) algorithm, utilizing interpolation of the spectral parameters between adjacent frames, is presented. Preliminary investigations, obtained in a CELP (code excited linear prediction) coder environment, indicated that time variability can improve speech quality in the case of long analysis frames (40 ms and above). A novel double-mode LTP structure was introduced to cope with the long start-up time of the conventional adaptive codebook. Segmental SNR measurements and listening tests showed that the double-mode LTP outperforms existing delayed decision tree coding algorithms with comparable complexity in the 10 ms subframe case. Delayed decision tree coding was also applied to the double-mode LTP, giving a further slight improvement of speech quality. The problem with computational complexity in the codebook search was shown to be solvable with a very small reduction of speech quality for 10 ms subframes, using a TRPE (transformed binary regular pulse excitation) codebook.<>