Opens in a new tab

AES E-Library

← Back to search

Conference Paper Open Access

Where Can Decay Losses Act? Characterizing EDR Supervision Across Training and Sampling in Diffusion RIR Synthesis

Authors: Baran, Mihai-Vlad; Kelly, Jack; Woszczyk, Wieslaw

AES Show 2026 Nashville · Paper 24 · October 2026

Abstract

Differentiable energy-decay losses have been shown to improve deterministic artificial reverberation algorithms, such as feedback delay networks. In this work, we investigate whether these benefits transfer to diffusion-based
room impulse response (RIR) generation. A v-prediction diffusion U-Net operating on complex spectrograms is trained on approximately 10,000 measured RIRs and evaluated on 384 unseen room impulse responses from six receiver placements. The proposed Energy Decay Relief (EDR) criterion is incorporated in three stages of the generation pipeline: direct training, post-training fine-tuning, and inference-time guidance. The baseline diffusion
model already reproduces the conventional room acoustic parameters (RT60, EDT, and C50) with errors below their perceptual thresholds, leaving the late reverberation tail as the primary remaining source of error. Although the
EDR criterion successfully influences the targeted decay characteristics in all three regimes, it does not improve the overall quality of the generated responses. Direct training produces only marginal improvements while slightly
increasing the reverberation time, fine-tuning improves the training objective by 1.1 dB without improving the generated responses, and inference-time guidance enhances the late reverberation tail only at the expense of
perceptually noticeable reductions in clarity. These conclusions are supported by paired statistical comparisons over the full hold-out dataset.

Details

Published in
AES Show 2026 Nashville
Paper number
24
Publication date
October 30, 2026
Session subject
Artificial Intelligence & Machine LearningRoom & Architectural Acoustics
Affiliation
McGill University (See document for exact affiliation information.)
Type
Conference Paper

Save to list