A Cold Diffusion Approach for Percussive Dereverberation

Researchers have introduced a novel cold diffusion framework designed specifically for the dereverberation of percussive and drum audio signals, addressing a significant gap in current audio processing technologies that predominantly focus on speech. This new approach models reverberation as a deterministic degradation process, transforming anechoic signals into reverberant ones to facilitate effective reversal. The study investigates two reverse-process parameterizations: Direct next-state prediction and Delta-normalized residual prediction, implemented using both UNet and diffusion Transformer backbones. Trained on curated datasets containing acoustic and electronic drum recordings with synthetic and real room impulse responses, the model demonstrates superior performance compared to existing score-based and conditional diffusion baselines. Extensive experiments on both in-domain and out-of-domain test sets confirm its effectiveness using signal-based and perceptual metrics tailored for percussive audio. This advancement holds significant potential for improving music production quality by enhancing the clarity of drum stems.