In this paper, we present a new speech dereverberation method using the kurtosis-maximization based on the voiced/unvoiced/silence (V/UV/S) classification. Since kurtosis of the UV/S sections are much smaller than V sections, adaptation of the dereverberation filter using these sections often results in slow and non-robust convergence, and, in turn, poor dereverberation. The proposed algorithm controls adaptation of the dereverberation filter using the results of V/UV/S classification, together with kurtosis measure of the input speech. For the selective control of adaptation, both hard decision and voice likelihood measure based on various features together with kurtosis were tried, and then, the step-size of the adaptive algorithm was varied according to various control strategies. The proposed algorithm provides better and more robust dereverberation performance than the conventional algorithm, which was confirmed through the experiments.
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.