High speed multi-stage code search algorithm in CELP

Abstract

An efficient multistage algorithm for code search in the code excited linear prediction (CELP) methods of speech coding is described. The algorithm requires binary clustering of the codebook into a fixed number of cells and inverse filtering of the incoming speech by the format and pitch filters to produce a residual error sequence (RES). It is demonstrated that the method could achieve a reduction in the number of operations from 40 to 50 times. It is shown that codebooks constructed from random walk stochastic sequences improve the average signal to noise ratio by 5-6 dB over codebooks constructed from Gaussian white noise sequences