Plug-and-Play WPE Guided by Deep Spectrum Estimation for Speech Dereverberation

Speech dereverberation aims to attenuate the effects of late-reverberant components. While the plug-and-play weighted prediction error (PnPWPE) method represents an innovative approach to dereverberation with exceptional performance, it can be further enhanced from several aspects. These include improving the accuracy of power spectral density (PSD) initialization to enhance energy normalization and optimizing the process of parameter selection. To address these areas for improvement, this paper introduces an enhanced PnPWPE framework. Within this framework, PSD initialization is supported by a deep neural network, and a dynamic strategy is implemented for parameter adjustment, eliminating the need for manual selection. Experimental findings validate the efficacy of the proposed method.