Absorbing Discrete Diffusion for Speech Enhancement

2026-02-25

AI Summary

This work addresses the challenge of speech enhancement under low signal-to-noise ratio (SNR) conditions by proposing a novel approach based on an absorbing discrete diffusion model. The method models the conditional distribution of clean speech codes in the expressive latent space of a neural audio codec in a non-autoregressive manner and introduces, for the first time, an absorbing discrete diffusion mechanism to the speech enhancement task. To this end, the authors design the RQDiT architecture, which integrates residual vector quantization (RQ) with a diffusion Transformer to effectively capture the hierarchical structure of RQ-encoded representations. Experimental results demonstrate that the proposed method achieves competitive non-intrusive objective scores on two benchmark datasets, particularly excelling in low-SNR scenarios and with very few sampling steps.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Computer Vision: Diffusion Models for Vision
• Cognitive Modeling & Cognitive Systems: Neural Spike Coding

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Data and user privacy-enhancing technologies for the Web
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Inspired by recent developments in neural speech coding and diffusion-based language modeling, we tackle speech enhancement by modeling the conditional distribution of clean speech codes given noisy speech codes using absorbing discrete diffusion. The proposed approach, which we call ADDSE, leverages both the expressive latent space of neural audio codecs and the non-autoregressive sampling procedure of diffusion models. To efficiently model the hierarchical structure of residual vector quantization codes, we propose RQDiT, which combines techniques from RQ-Transformer and diffusion Transformers for non-autoregressive modeling. Results show competitive performance in terms of non-intrusive objective metrics on two datasets, especially at low signal-to-noise ratios and with few sampling steps. Code and audio examples are available online.

Problem

• speech enhancement
• discrete diffusion
• neural speech coding
• conditional distribution
• noisy speech

Innovation

• absorbing discrete diffusion
• speech enhancement
• neural audio codec
• non-autoregressive modeling
• RQDiT