BAYESIAN CHANNEL EQUALIZATION AND ROBUST FEATURES FOR SPEECH RECOGNITION

The use of a speech recognition system with telephone channel environm ents, or different microphones, requires channel equalisation. In spee ch recognition, the speech model provides a bank of statistical inform ation that can be used in the channel identification and equalisation process. The authors consider HMM-based channel equalisation, and pres ent results demonstrating that substantial improvement can be obtained through the equalisation process. An alternative method, for speech r ecognition, is to use a feature set which is more robust to channel di stortion. Channel distortions result in an amplutude tilt of the speec h cepstrum, and therefore differential cepstral features provide a mea sure of immunity to channel distortions. In particular the cepstral-ti me feature matrix, in addition to providing a framework for representi ng speech dynamics, call be made robust to channel distortions. The au thors present results demonstrating that a major advantage of cepstral -time matrices is their channel insensitive character.