A Bilinear Framework For Adaptive Speech Dereverberation Combining Beamforming And Linear Prediction

Speech dereverberation algorithms based on multichannel linear prediction (MCLP) are effective under various acoustic conditions. This paper proposes a bilinear form for the MCLP based dereverberation, where the MCLP filter is expressed as a Kronecker product of a spatial filter and a temporal filter. Then, a recursive least-squares (RLS)-based algorithm is derived for adaptive speech dereverberation. Compared with the original MCLP-based adaptive algorithm, the advantages of the proposed method are twofold: (1) the computational complexity is significantly reduced and is more suitable for dynamic scenarios, since fewer parameters have to be estimated per signal-block observation; and (2) it is more robust to noise by optimizing the spatial filter as a weighted minimum power distortionless response (wMPDR) beamformer. Simulation results validate the advantages of the proposed algorithm.