A Cold Diffusion Approach for Percussive Dereverberation

Dimos Makris, András Barják, Maximos Kaliakatsos-PapakostasarXiv 2026 · arXiv

TL;DR

This paper introduces the first learning-based framework specifically designed for blind dereverberation of percussive and drum signals. By leveraging a "cold diffusion" approach, the model treats reverberation as a deterministic degradation process rather than stochastic noise, significantly outperforming traditional speech-centric diffusion baselines in preserving sharp transients and rhythmic clarity.

What Problem Does This Paper Address?

While audio dereverberation has advanced rapidly thanks to deep generative models, existing solutions are almost exclusively optimized for speech. Percussive signals (like drum kits) present entirely different challenges: they are dominated by sharp, fast-decaying transients and lack continuous harmonic or phonetic structures to guide a model.

When reverberation is applied to drums, it smears these sharp onsets and redistributes energy into the tail, blurring the rhythmic timing essential for music production. Standard speech-enhancement models and evaluation metrics (like PESQ or STOI) fail to handle these dense, overlapping resonances. This paper tackles the specific task of Drums Dereverberation on stereo downmixes (stems), aiming to cleanly remove room acoustics while keeping the drum hits crisp and punchy.

Methods

Instead of using standard diffusion models that add and remove random Gaussian noise, the authors adopt Cold Diffusion. Here, the forward process is a deterministic degradation operator that progressively blends an anechoic (clean) spectrogram into its reverberant counterpart using a cosine-squared mixing schedule.

Illustration of the forward and reverse cold diffusion process on a drum spectrogram.

The authors explore two ways to parameterize the reverse (restoration) process:

  1. Direct Prediction: The network directly predicts the next, less-reverberant state.
  2. $\Delta$-Normalized Residual Prediction: The network predicts a step-size-normalized "velocity" vector representing the difference between consecutive states, which helps stabilize the iterative generation.

They implement this framework using two distinct neural backbones: a modified UNet (with a specialized $9 \times 1$ input kernel tailored to transient spectra) and a Diffusion Transformer (DiT). The training objective combines a multi-resolution STFT loss (capturing both short-window transients and long-window decay tails) with a time-domain waveform loss.

Datasets

To train the models, the authors compiled a 38-hour paired anechoic-reverberant dataset from:

  • MUSDB18-HQ: Real acoustic drum recordings.
  • Groove MIDI Dataset (GMD): Electronically rendered drum performances.

Crucially, an experienced audio engineer manually curated the source files to ensure they were perceptually "dry" (lacking pre-existing room tails or heavy ambience). Reverberant versions were generated by convolving these dry stems with both synthetic room impulse responses (RIRs) via pyroomacoustics and real-world RIRs from the OpenAIR database.

For evaluation, they built a fully out-of-domain test set using dry drum stems from MoisesDB convolved with real physical room responses from the ACE Challenge Workshop, ensuring a rigorous test of the model's ability to generalize.

Key Insights and Ideas

  • Deterministic > Stochastic for Reverb: Standard diffusion assumes noise is additive and random. Cold diffusion treats reverberation as what it actually is—a structured, deterministic distortion—making it far more effective at learning the inverse mapping.
  • The Power of Step-Normalization: The $\Delta$-normalized residual parameterization proved vital, especially for the Transformer backbone. By constraining each step to a schedule-aware update, it mitigates the "exposure bias" (error accumulation) common in iterative sampling chains.
  • Backbone Inductive Biases: The UNet consistently outperformed the Diffusion Transformer (DiT) on transient-sensitive metrics. The authors suggest that the structural priors of convolutional networks remain highly advantageous for processing the sharp, localized energy of percussive audio.

Key Results

The proposed cold diffusion models were compared against strong diffusion baselines originally designed for speech: SGMSE+ and CDiffuSE.

  • In-Domain Performance: The Cold UNet $\Delta$-norm model achieved an impressive 11.09 dB Scale-Invariant Signal-to-Distortion Ratio improvement (SI-SDRi), vastly outperforming SGMSE+ (4.06 dB) and CDiffuSE (2.77 dB). It also demonstrated superior envelope correlation (ENV) and transient-to-tail energy ratio (TTER).
  • Out-of-Domain Generalization: When tested on entirely unseen drums and rooms, the speech baselines collapsed (SI-SDRi dropped to 2.01 dB and 0.17 dB). Meanwhile, the Cold UNet $\Delta$-norm retained a robust 7.52 dB SI-SDRi, proving its real-world viability.
Qualitative spectrogram comparison showing how the cold diffusion models cleanly remove late reverberation compared to the speech baseline.

Limitations and Future Work

While the model excels at removing physical room reflections, the authors note a few limitations:

  • Low-Frequency Attenuation: In extreme, highly reverberant unseen conditions, the UNet occasionally over-attenuates low-frequency content (e.g., thinning out the kick drum).
  • Physical vs. Production Reverb: The current model is trained on physical room acoustics (RIRs). It does not yet account for artificial, non-physical production effects like plate or spring reverbs, which are common in commercial music.

Future work will focus on expanding the dataset to include these production-style algorithmic reverbs, exploring variable-step inference schedules to reduce computational overhead, and incorporating transient-sensitive evaluation metrics directly into the training loss functions.

Summary updated

Abstract

Most recent advances in audio dereverberation focus almost exclusively on speech, leaving percussive and drum signals largely unexplored despite their importance in music production. Percussive dereverberation poses distinct challenges due to sharp transients and dense temporal structure. In this work, we propose a cold diffusion framework for dereverberating stereo drum stems (downmixes), modeling reverberation as a deterministic degradation process that progressively transforms anechoic signals into reverberant ones. We investigate two reverse-process parameterizations, Direct (next-state) and a Delta-normalized residual (velocity-style) prediction, and implement the framework using both a UNet and a diffusion Transformer backbone. The models are trained and evaluated on curated datasets comprising both acoustic and electronic drum recordings, with reverberation generated using a combination of synthetic and real room impulse responses. Extensive experiments on in-domain and fully out-of-domain test sets demonstrate that the proposed method consistently outperforms strong score-based and conditional diffusion baselines, evaluated using signal-based and perceptual metrics tailored to percussive audio.
BibTeX
@misc{makris2026cold,
  author = {Dimos Makris and András Barják and Maximos Kaliakatsos-Papakostas},
  title = {{A Cold Diffusion Approach for Percussive Dereverberation}},
  year = {2026},
  eprint = {2605.10256},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.10256},
}

Loading related content...

Figure