Corrective forcing improves speech quality in generative enhancement models
Corrective Forcing: Unified Post-Training for Diffusions and Flows in Generative Speech Enhancement
Machine LearningArtificial Intelligence
Summary
Speech enhancement models that generate clean audio can make errors when they try to fix noisy speech step-by-step, because they are trained differently than they are used in practice. The authors introduce a method called Corrective Forcing that lets these models learn from their own mistakes after training, improving their ability to produce clearer speech. This method works for different kinds of generative models and helps them make better predictions even when the process is stopped early or takes more steps. Overall, it results in more natural-sounding and accurate speech enhancement.
What this means in practice
- •For audio software developers: Improve speech enhancement features in apps by reducing noise and enhancing clarity with generative models trained using Corrective Forcing.
- •For voice communication engineers: Enhance live or recorded voice call quality by applying models corrected with Corrective Forcing to handle varied noise conditions robustly.
Authors
Qing Yao, Lijian Gao, Qirong Mao
Abstract
Diffusion and flow models, as promising generative paradigms for speech enhancement, face a training--inference mismatch: training uses analytical path states, whereas inference recursively evaluates models on self-generated rollout states along discretized sampling trajectories. This mismatch causes prediction and discretization errors to accumulate. To address it, we introduce Corrective Forcing (CoF), a post-training paradigm that forces diffusion and flow models to learn from self-generated rollouts and correct their predictions. CoF corrects clean-speech predictions on rollout states toward the ground truth under dynamic sampling schedules, exposing the model to varying inference conditions. It further regularizes local evolution using locally corrected counterfactual transitions as references for factual transitions. By expressing model outputs through a shared clean-speech prediction parameterization, CoF applies the same post-training objective across diffusion and flow formulations. Experiments with SB-VE and OT-CFM demonstrate improvements in perceptual quality and reconstruction fidelity, together with robust performance across different numbers of sampling steps.