Frame-synchronous adaptation of cepstrum by linear regression

We propose using hidden Markov models (HMMs) associated with the cepstrum coefficients to remove disturbances that degrade the speech recognition process. In order to perform this task in an online manner, we use the MUltipath Stochastic Equalization (MUSE) framework. This method allows one to process data at the frame level. Two equalization functions are examined: bias removal and linear regression. Recognition experiments carried out on both PSTN and GSM networks show the efficiency of the proposed method: thanks to MUSE, with a model trained on PSTN recorded digits, the error rate on both PSTN and GSM recorded digits can be reduced by 19% with bias subtraction and by 36% with linear regression. Similar results obtained on another vocabulary are also presented.