Sparse Linear Prediction-based Dereverberation for Signal Enhancement in Distant Speaker Verification

In this paper we investigate several sparse dereverberation methods which provide signal enhancement in reverberant environments with the aim to apply them as a preprocessing step for the distant speaker verification (SV) task. First, we present multichannel linear prediction (LP) based techniques which promote sparsity of the dereverberated speech, whose performance has never been verified in the context of speaker recognition. In particular, we describe two existing sparse LP-based methods and present a novel LP-based method in which speech sparsity is enforced by adopting the so-called split Bregman approach. We then study the performance of both sparse and nonsparse dereverberation approaches for signal enhancement, and investigate the gain offered by these methods when applied as a preprocessing step for two different speaker verification systems based on DNN-based speaker embedding extraction. The results of performed experiments indicate that the proposed sparse approach and one of the existing methods consistently achieve significant improvements in distant speaker verification in reverberant environments, and that the SV results are well in line with signal enhancement achieved by the compared techniques.