A frequency-selecting multichannel equalization method for speech dereverberation

A frequency-selecting approach is proposed for robust multichannel dereverberation in the presence of system identification errors (SIEs), based on subband equalization of acoustic channels. Same as the full-band counterpart, performance of the subband multichannel equalizer (SB-MCEQ) degrades in the presence of SIEs. Firstly performance of the SB-MCEQ in different subbands is investigated. Statistical results show that the performance of SB-MCEQ degrades severely only in a few subbands, and they will be called ‘bad’ subbands. Then a frequency-selecting multichannel equalization (FS-MCEQ) algorithm is proposed, which abandons the subband signals of the ‘bad’ subbands or leaves them unprocessed at reconstruction of the full-band signal from the subband signals. Simulation results confirm that the FS-MCEQ algorithm outperforms the SB-MCEQ in the presence of SIEs.