New Journal on Buddy: Unsupervised Dereverberation Using Diffusion

6 June 2025, by David Mosteller

We are delighted to share that our extended journal article on "Buddy" has been accepted for publication in IEEE Transactions on Audio, Speech and Language Processing and is now available as early access:

Jean-Marie Lemercier, Eloi Moliner, Simon Welker, Vesa Välimäki, Timo Gerkmann, "Unsupervised Blind Joint Dereverberation and Room Acoustics Estimation with Diffusion Models", IEEE Trans. Audio, Speech, Language Proc., accepted. [doi] [arxiv] [audio] [code]

This follows our earlier work, "Buddy: Single-Channel Blind Unsupervised Dereverberation with Diffusion Models"—co-authored with Eloi Moliner, Jean-Marie Lemercier, Simon Welker, and Vesa Välimäki—which received the Best Student Paper Award at IWAENC 2024 in Aalborg.

The TASLP journal article offers a significantly broader and deeper evaluation of the BUDDy algorithm, including:

- Subjective listening tests
- Application to singing voice dereverberation
- Extensive ablation studies
- Benchmarking against state-of-the-art RIR estimators
- Real-world robustness analysis

These additions further demonstrate the versatility and reliability of our fully blind, unsupervised dereverberation approach based on diffusion models. Many thanks to my fantastic co-authors and everyone who supported this work!

Do not miss our listening examples on our webpages.