Reduced complexity CELP coder

Code excited linear predictor (CELP) coders enable speech coding with good quality at bit rates as low as 4 kbps. Several methods have been developed to lessen the computational task. The authors present a multistage method that preserves the qualities of the CELP vocoder with reduced complexity. The basic idea is to find a subset S₁ of candidate codewords by approximating the synthetic signal energy. This approximation is obtained by restricting the frequency range and subsampling. A full exact search is done on this subset. The procedure can be iterated. Typically, for an initial codebook of 1024 words and a single stage of subsampling by a factor equal to q ₁=5, the size of the subset S₁ is 70 and the calculation gain is 9. This performance can be improved up to a calculation gain of 13 by introducing several stages of subsampling q₁=5 and q₂=2