A recursive generalized sidelobe canceler for multichannel blind speech dereverberation

In this paper, we propose a generalized sidelobe canceler for multichannel blind speech dereverberation, which relies on recursive estimation of posterior distributions on the unknown acoustic channels and the adaptive interference canceler (AIC). Contrary to conventional design approaches where a fixed beamformer is employed, we consider a marginalized maximum-likelihood equalizer that is driven by the channel posterior estimator. It is shown that the first moment of the inferred channel posterior can also serve as a representation of an adaptive blocking matrix (ABM). Using the output of the blocking matrix, we estimate the AIC posterior to minimize the residual reverberation in the equalized signal. We demonstrate the efficacy of our approach by evaluating the algorithm in different degrees of observation noise and varying reverberation times.