Channel and speaker adaptation techniques for robust speech recognition

Acoustical mismatches between training and testing environments of HMM-based speech recognizer often cause sever degradations in recognition performance.The mismatch is mainly caused by noise,changes of the channel through which the speech signal is transmit-ted and differences of speakers.This paper addresses the problem of changes of the recording channel and variations of speakers.Three udaptation methods are presented to deal with the problem,i.e.,the adaptation via VQ prototype modification,the adaptation via HMM parame-ters modification and the canonical correlation based compensation method (short for CCBC).Experimental results have shown that all the three techniques can make our speaker-independent recognition system robust to channels and speakers.Among the three techniques,the CCBC has the best performance and it can be used as a unified approach to cope with mismatch caused both by noise,by differences of channels and by variations of speakers.