Compensating additive noise and CS-CELP distortion in speech recognition using stochastic weighted Viterbi algorithm

A solution to the problem of speech recognition with signals corrupted by additive noise and CS-CELP coders is presented. The additive noise and the coding distortion are cancelled according to the following scheme: first, the pdf of the clean coded-decoded speech is estimated with an additive noise model; secondly, the pdf of the clean uncoded signal is also estimated with a coding distortion model; finally, the hidden Markov model is compensated using the expected value of observation pdf in the context of the stochastic weighted Viterbi algorithm.