Distant-talking speech recognition using multi-channel LMS and multiple-step linear prediction

Previously, dereverberation methods based on generalized spectral subtraction (GSS) using multi-channel least mean squares (MCLMS) and multiple-step linear prediction (MSLP) have been proposed. Both methods have in common to estimate the late reverberation characteristics blindly, to suppress the late reverberation by spectral subtraction. Speech recognition performances of both methods are changing according to length of late reverberation to be estimated. In this paper, we investigated effect of estimated length of late reverberation on distant-talking speech recognition. Moreover, we proposed method to combine MCLMS and MSLP. As a result, MCLMS-based dereverberation method is effective to reduce in the long reverberation with approximately 200 ms and MSLP dereverberation is effective for the short reverberation with approximately 100 ms. The proposed method of “MSLP+MCLMS” (that is, MCLMS is applied after MSLP) outperformed than all other dereverberation methods.