Corrective Forcing Reduces Error Accumulation in Speech Diffusion
September 22, 2026
Corrective Forcing (CoF) is a post-training paradigm for speech enhancement that forces diffusion and flow models to learn from self-generated rollout states. This addresses the training-inference mismatch by regularizing local evolution using locally corrected counterfactual transitions.
HOW THIS AFFECTS YOU
●
researcherYou can mitigate discretization and prediction errors in generative speech models through post-training.