A speech dereverberation method based on the MTF concept using adaptive time-frequency divisions

We previously proposed a speech dereverberation method based on the MTF concept. In the proposed model, the power envelopes and carriers are decomposed from a reverberant speech signal using a constant-bandwidth filterbank and then are restored in each respective channel using power envelope inverse filtering and a carrier regeneration method, respectively. In this paper, to develop the model, we improve it through adaptive time-division and frequency-division processing using a reconstructed filterbank depending on each speech signal. We have carried out 15,000 simulations of dereverberation for reverberant speech signals to evaluate the improved model. We found that the improved model can restore a speech signal from reverberant speech and can improve upon the restoration accuracy of the previous model - that is, reduce the average log-spectrum distortion - by about 1 dB compared to results obtained using our previous model and the time- and/or frequency-division model.