Unsupervised Blind Joint Dereverberation and Room Acoustics Estimation with Diffusion Models

TL;DR

This work introduces BUDDy, an unsupervised, diffusion-model–based framework for blind single-channel dereverberation and room impulse response (RIR) estimation. It couples a diffusion-based prior on the anechoic speech with a differentiable, parametric subband RIR model optimized via an EM-like process, enabling joint estimation of $x_0$ and $h$ without paired data. The method demonstrates strong performance in both speech and singing-voice dereverberation, robust generalization to mismatched acoustic conditions, and competitive RIR estimation compared with supervised baselines, albeit with higher computational cost. By jointly optimizing the RIR operator and performing diffusion posterior sampling, BUDDy achieves robust dereverberation and accurate room acoustics characterization, offering practical advantages in real-world, unseen environments.

Abstract

This paper presents an unsupervised method for single-channel blind dereverberation and room impulse response (RIR) estimation, called BUDDy. The algorithm is rooted in Bayesian posterior sampling: it combines a likelihood model enforcing fidelity to the reverberant measurement, and an anechoic speech prior implemented by an unconditional diffusion model. We design a parametric filter representing the RIR, with exponential decay for each frequency subband. Room acoustics estimation and speech dereverberation are jointly carried out, as the filter parameters are iteratively estimated and the speech utterance refined along the reverse diffusion trajectory. In a blind scenario where the RIR is unknown, BUDDy successfully performs speech dereverberation in various acoustic scenarios, significantly outperforming other blind unsupervised baselines. Unlike supervised methods, which often struggle to generalize, BUDDy seamlessly adapts to different acoustic conditions. This paper extends our previous work by offering new experimental results and insights into the algorithm's versatility. We demonstrate the robustness of our proposed method to new acoustic and speaker conditions, as well as its adaptability to high-resolution singing voice dereverberation, using both instrumental metrics and subjective listening evaluation. We study BUDDy's performance for RIR estimation and observe it surpasses a state-of-the-art supervised DNN-based estimator on mismatched acoustic conditions. Finally, we investigate the sensitivity of informed dereverberation methods to RIR estimation errors, thereby motivating the joint acoustic estimation and dereverberation design. Audio examples and code can be found online.

Figures (5)

• Figure 1: BUDDy: joint optimization alternating between RIR estimation and posterior sampling for speech reconstruction moliner2024buddy.
• Figure 2: Listening test results on VCTK-RealReverb and VCTK-SimulatedReverb. The boxplot shows first quartile, median, and third quartile.
• Figure 3: Listening test results on singing voice datasets NHSS-RealReverb and NHSS-SimulatedReverb. The boxplot shows first quartile, median, and third quartile.
• Figure 4: Robustness of informed dereverberation approachess with respect to normally distributed errors in the RIR.
• Figure 5: RIR estimation metrics for each octave and full-band on the reverberant VCTK dataset. The violin plots show the distribution and the median. Lower is better. FiNS steinmetz2021filtered is trained in a supervision fashion whereas BUDDy is unsupervised.