Distributed Weighted Prediction Error for Speech Dereverberation with Regularization by Denoising

Speech dereverberation addresses the degradation of speech quality caused by late reverberation. Although the weighted prediction error (WPE) method has demonstrated superior performance in mitigating reverberation, its centralized architecture results in substantial computational and communication overhead, particularly in distributed settings where each spatially separated node is equipped with a microphone array. This paper first formulates a novel distributed WPE optimization problem that fits into this network scenario. To further enhance the optimization process, we propose to integrate data-driven speech priors into the framework via a plug-and-play strategy. Hence, the proposed framework not only reduces the computation and communication complexity at individual nodes through effective inter-node collaboration but also improves performance under challenging acoustic conditions. Experimental evaluations confirm the framework’s effectiveness in both noise-free and noisy distributed scenarios.