BUDDy: Single-Channel Blind Unsupervised Dereverberation with Diffusion Models

Generative models represent another category of dereverberation algorithms aiming to learn the distribution In this paper, we present an unsupervised single-channel method for of anechoic speech conditioned on reverberant input. Some joint blind dereverberation and room impulse response estimation, blind supervised methods using generative models such as diffusion based on posterior sampling with diffusion models. We parameterize models [12, 13] have been recently proposed [14, 15]. However, supervised the reverberation operator using a filter with exponential decay approaches struggle with limited generalization to diverse for each frequency subband, and iteratively estimate the corresponding acoustic conditions due to the scarcity and variability of available parameters as the speech utterance gets refined along the reverse RIR data. Unsupervised approaches offer the potential to circumvent diffusion trajectory. A measurement consistency criterion enforces such limitations as they do not require paired anechoic/reverberant the fidelity of the generated speech with the reverberant measurement, data. This paper builds upon prior work [16], which proposed an while an unconditional diffusion model implements a strong unsupervised method for informed single-channel dereverberation prior for clean speech generation. Without any knowledge of the based on diffusion posterior sampling. The previous study showed room impulse response nor any coupled reverberant-anechoic data, the potential of leveraging diffusion models as a strong clean speech we can successfully perform dereverberation in various acoustic scenarios.