Speech Dereverberation with Deconvolution Regularized by Denoising

Deconvolution-based speech dereverberation continues to present challenges due to the difficulties in accurately acquiring Room Impulse Responses (RIRs) and the inherently ill-conditioned nature of deconvolution. Despite advancements in RIR measurement and estimation, substantial room for improvement remains in addressing the latter challenge. This paper proposes a novel prior-driven dereverberation framework utilizing Regularization by Denoising (RED) to incorporate data priors into the deconvolution process, thereby addressing this persistent challenge. Specifically, we formulate the dereverberation process via an optimization problem with the additional regularizer and the Half Quadratic Splitting (HQS) strategy is then utilized to solve the optimization problem. Experimental validation conducted on both the RIR simulation platform pyroomacoustics and the realistic acoustics platform SoundSpaces demonstrates the efficacy of our framework, even in the presence of environmental noise and RIR errors.