Robust Speech Dereverberation Based on WPE and Deep Learning

Reverberation has a considerable impacts on speech quality and intelligibility. Weighted prediction error (WPE) employs a linear regression model to estimate late reverberation and then cancel it. The key point of the WPE is to estimate the power spectrum of the early speech. However, its estimation relies on an iterative procedure with high computational complexity. Another problem is that the WPE has a noise-free assumption. So, the performance degrades in noisy conditions. To address these problems, we propose an algorithm for speech dereverberation in the presence of background noise, in which deep learning is integrated into the WPE method. Specifically, we employ a neural network to predict the power spectral density (PSD) of early speech and a binary mask which distinguishes target speech from background noise. To alleviate the noise impact on estimation of echo path, a dual-filter strategy is adopted to model the echo paths of target speech and background noise individually. Experimental results show that the proposed method significantly improves speech quality in noisy environments.