Neural-Network Supervised Maximum Likelihood-based on-line Dereverberation

In this paper, a new online multiple-input multipleoutput (MIMO) approach based on Maximum Likelihood (ML) in subband-domain for dereverberation is proposed. Multichannel linear prediction filters are estimated to blindly shorten the Room Impulse Responses (RIRs) between a set of unknown number of sources and a microphone array. The adaptive filter is updated using a modified weighted recursive Least Squares (RLS). To speed up convergence and minimize the influence of noise, the adaptive algorithm is supervised by a trained Deep Neural Network (DNN) which predicts the source dominance. In our experiments, it is proved that the proposed method can largely reduce the effect of reverberation in high non-stationary noisy conditions and sensibly improve automatic speech recognition performance in far-field and high reverberation.