Hierarchical Graph-coding Diffusion Model with Adaptive Information Bottleneck for Multichannel Speech Enhancement

ICLR 2026 Conference Submission25493 Authors

Keywords: hierarchical graph-coding, diffusion model, layer modulation, adaptive information bottleneck, multichannel speech enhancement

Abstract: Diffusion models have achieved strong performance in multichannel speech enhancement, especially in unseen noisy scenarios. However, most existing diffusion method rely on globally consistent guidance applied either to the output or uniformly across denoiser layers, which fails to provide layer-specific adaptation and introduces redundancy, thereby constraining denoising performance.To address these challenges,we propose a novel hierarchical graph-coding diffusion model with adaptive information bottleneck (HG-Diff-IB) for multichannel speech enhancement. Specifically, we introduce a hierarchical alignment method to align graph-coding with the denoiser at different depths, together with a layer-wise graph-coding modulation mechanism that injects graph information into intermediate features, enabling layer-specific guidance of diffusion feature distributions. Furthermore, we introduce an adaptive information bottleneck that dynamically adjusts the feature compression according to the estimated SNR, effectively balancing noise suppression and target feature preservation. Experimental results demonstrate that our proposed method outperforms baselines in various evaluation metrics.

Primary Area: generative models

Submission Number: 25493