Blind Deconvolution for Multi-microphone Speech Dereverberation: Application to Asr in Reverberant Environments

In this paper, a deterministic time-domain algorithm for multichannel blind deconvolution is presented. The proposed algorithm assumes that a source signal is measured by several sensors after propagating through finite impulse response channels and being corrupted by additive noise. An estimate of the source signal is obtained by first estimating the channel impulse responses in a Least Squares sense and by next computing exact deconvolution filters in a Minimum Norm sense. In moderate noise conditions, the proposed algorithm is shown to perform satisfactorily. More especially, we apply this technique for dereverberating speech recorded in a reverberant enclosure via distant microphones. We report results demonstrating that the proposed dereverberation technique, when used as a pre-processing block, improves significantly the performance of an Automatic Speech Recognition (ASR) system operated in a reverberant enclosure.