🔗 Project page
🔗 Paper DriftSE
DriftSE is a speech enhancement method that reformulates denoising as a distribution transport problem.
Instead of iterative denoising or direct regression, it learns a drifting field that moves noisy speech distributions toward clean speech.
- Speech enhancement = distribution evolution
- Learn a drift field to guide noisy → clean speech
- Enables one-step inference
DriftSE replaces iterative denoising with a learned continuous transformation of distributions, achieving fast and effective speech enhancement in a single step.