Speech Enhancement Based on Drifting Models

📰 ArXiv cs.AI

Learn to enhance speech quality using Drifting Models, a novel generative framework that formulates denoising as an equilibrium problem

advanced Published 28 Apr 2026

Action Steps

Formulate denoising as an equilibrium problem using Drifting Models
Evolve the pushforward distribution of a mapping function to match the clean speech distribution
Implement a Drifting Field to guide samples toward the clean speech distribution
Train the Drifting Field using a suitable loss function and optimization algorithm
Evaluate the performance of the DriftSE framework using metrics such as SNR and PESQ

Who Needs to Know This

Audio engineers and researchers working on speech enhancement tasks can benefit from this framework to improve speech quality in noisy environments

Key Insight

💡 Drifting Models can be used to formulate denoising as an equilibrium problem, allowing for one-step inference and improved speech quality

Key Takeaways

Learn to enhance speech quality using Drifting Models, a novel generative framework that formulates denoising as an equilibrium problem

Full Article

Title: Speech Enhancement Based on Drifting Models

Abstract:
arXiv:2604.24199v1 Announce Type: cross Abstract: We propose Speech Enhancement based on Drifting Models (DriftSE), a novel generative framework that formulates denoising as an equilibrium problem. Rather than relying on iterative sampling, DriftSE natively achieves one-step inference by evolving the pushforward distribution of a mapping function to directly match the clean speech distribution. This evolution is driven by a Drifting Field, a learned correction vector that guides samples toward t

Read full paper → ← Back to Reads