Distributed Weighted Prediction Error with Node Selection Strategy for Speech Dereverberation

Speech dereverberation seeks to remove late reflections that smear the temporal and spectral structure of far-field recordings. Although the weighted prediction error (WPE) method has achieved promising performance, its centralized architecture incurs prohibitive computational and communication overhead in distributed scenarios where microphones are spatially dispersed and connected to resource-constrained processors. In this paper, we first formulate a distributed WPE optimization that incorporates an envelope-variance (EV)–based node selection module to exclude low-quality nodes and focus cooperation on the reliable observations. We then enhance the per-node optimization by integrating deep speech priors via Regularization-by-Denoising (RED), leveraging a pretrained deep neural network (DNN) denoiser as a proximal operator. Thus, the proposed framework not only balances computational load and reduces inter-node bandwidth by exchanging compressed signal, but also yields substantial dereverberation gains in challenging acoustic and noisy environments. Experiments under both simulated and real-world conditions confirm the superiority of the proposed method.