Posterior Transition Modeling for Unsupervised Diffusion-Based Speech Enhancement

2025-07-03

AI Summary

Unsupervised diffusion-based speech enhancement suffers from reliance on hyperparameter tuning and inaccurate approximation of likelihood scores. Method: We propose two hyperparameter-free algorithms: (i) integrating diffusion priors with an observation model, and (ii) defining the diffusion process directly on noisy speech—both explicitly modeling the conditional reverse transition distribution. Crucially, we introduce the posterior transition modeling paradigm, eliminating hyperparameter dependence and enabling analytical computation of the exact likelihood score. Contribution/Results: Our approach ensures theoretical rigor and engineering practicality. Evaluated on WSJ0-QUT and VoiceBank-DEMAND, it consistently outperforms both supervised and unsupervised baselines, yielding significant improvements in PESQ (+1.2–1.8), STOI (+3.5–5.1%), and enhanced robustness to domain shift.

Technology Category

• Machine Learning: Unsupervised & Self-Supervised Learning
• Natural Language Processing: Speech
• Computer Vision: Diffusion Models for Vision

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

We explore unsupervised speech enhancement using diffusion models as expressive generative priors for clean speech. Existing approaches guide the reverse diffusion process using noisy speech through an approximate, noise-perturbed likelihood score, combined with the unconditional score via a trade-off hyperparameter. In this work, we propose two alternative algorithms that directly model the conditional reverse transition distribution of diffusion states. The first method integrates the diffusion prior with the observation model in a principled way, removing the need for hyperparameter tuning. The second defines a diffusion process over the noisy speech itself, yielding a fully tractable and exact likelihood score. Experiments on the WSJ0-QUT and VoiceBank-DEMAND datasets demonstrate improved enhancement metrics and greater robustness to domain shifts compared to both supervised and unsupervised baselines.

Problem

• Unsupervised speech enhancement using diffusion models
• Improving reverse diffusion with exact likelihood scores
• Enhancing robustness to domain shifts in speech

Innovation

• Models conditional reverse transition distribution directly
• Integrates diffusion prior with observation model
• Defines diffusion process over noisy speech