Speech dereverberation using weighted prediction error with prior learnt from data

Speech dereverberation aims to mitigate the impact of late-reverberant components. As a typical approach to dereverberation, the weighted prediction error (WPE) method has shown its superior performance, however it is still possible to further improve its performance and robustness by incorporating sophisticated speech priors. Recent research demonstrates that the integration of physics-based and data-driven methods can improve the performance of various signal processing tasks while maintaining the interpretability of the problem solving process. Motivated by the relevant progress, this paper presents a novel dereverberation framework that incorporates the data-driven method for speech prior capturing for WPE. The plug-and-play strategy (PnP), specifically the regularization by denoising (RED) strategy, is used to incorporate speech prior information during the alternating direction method of multipliers (ADMM) solving iterations by plugging in a pre-trained speech denoiser. Experimental results demonstrate the effectiveness of the proposed method.