A multi-microphone sub-band selection method has been proposed for dereverberation with the assumption that the transfer function between the sound source and each micro-phone can be calibrated. In this paper, the dereverberation system is implemented for a speaker dependent word recognizer as a preprocessor. The degradation of the recog-nition performance was investigated under three reverberation conditions. The prepro-cessor is evaluated by the recognition rate of using different numbers of microphones and sub-bands. It is found, when the distance between the sound source and the microphone is about 2.65m and the reverberation times of the enclosure are 0.55, 0.68, 0.86s, the recognition rates of the reverberated speech decreased by 12.8, 13.6, 14.5%, respectively, compared with the original clean speech. The results showed that the recognition rates were improved by 11.3, 12.6, 12.3%, respectively, by introducing the dereverberation preprocessor.
View full abstract