A unified adaptation approach for robust speech recognition

An acoustical mismatch between the training and testing conditions of hidden Markov model (HMM)-based speech recognition systems often causes a severe degradation in the recognition performance. The mismatch is mainly caused by noise, changes of recording channel and differences of speakers. In previous work^[1], we have used a canonical correlation based compensation scheme successfully to make our recognizer robust to the noise. In this paper, the CCBC technique is extended to include effects of channel and speakers. Experiment has been performed and the result shows that the CCBC is able to drastically reduce the word error rate of our speaker-independent recognition system. Compared with some conventional adaptation methods, such as Cepstral Mean Removal, RASTA and Lin-Log RASTA, the CCBC has superior performance and can be used as a unified adaptation approach for robust speech recognition.