Improved CELP Coding Using a Fully Adaptive Excitation Codebook

Most Code-Excited-Linear-Prediction (CELP) schemes use an excitation codebook containing stochastic excitation vectors, i.e. vectors consisting of random numbers [2]. Moreover, the codebook itself is a random but fixed collection of such vectors. The only speech specific information in the excitation arises from the memory of the long-term predictor. Even with an adaptive postfilter [1], CELP coded speech shows at data rates below 8 kb/s significant quality degradations in voiced parts of the speech signal, i.e. at long vowels and vowel transitions. Obviously, the variety of excitations generated by a stochastic codebook together with the adaptive codebook, represented by the long-term predictor, is not sufficient. Here we propose a CELP scheme, called ACELP, where the stochastic excitation codebook is replaced by an adaptive excitation codebook [3,4].