Corrective Forcing: Unified Post-Training for Diffusions and Flows in Generative Speech Enhancement

Diffusion and flow models have emerged as promising generative paradigms for speech enhancement. However, they suffer from a fundamental training–inference mismatch: during training, models are optimized on analytical path states, whereas at inference they recursively evaluate themselves on self-generated rollout states along discretized sampling trajectories. This discrepancy causes prediction and discretization errors to accumulate over the sampling process, degrading output quality.

The paper introduces **Corrective Forcing (CoF)**, a post-training paradigm that addresses this mismatch by forcing diffusion and flow models to learn from their own self-generated rollouts and correct their predictions. CoF is designed to be applicable across both diffusion and flow formulations through a shared clean-speech prediction parameterization, making it a unified post-training objective. The authors evaluate CoF with two representative instantiations: **SB-VE** (a variance-exploding diffusion model) and **OT-CFM** (optimal-transport conditional flow matching), demonstrating improvements in perceptual quality and reconstruction fidelity, along with robust performance across different numbe

The authors conduct experiments with two instantiations: **SB-VE** and **OT-CFM**. The results demonstrate improvements in perceptual quality and reconstruction fidelity. Specifically, CoF yields better performance compared to baseline models without post-training, as measured by standard speech enhancement metrics (though the abstract does not list specific metric values).

A key finding is that CoF provides robust performance across different numbers of sampling steps. This is particularly important because diffusion and flow models often degrade when the number of sampling steps is reduced for faster inference. CoF's ability to maintain quality across step counts suggests that it effectively mitigates the accumulation of discretization errors.

The experiments validate the unified nature of CoF: the same post-training objective improves both diffusion (SB-VE) and flow (OT-CFM) models, confirming that the shared clean-speech prediction parameterization enables cross-formulation applicability.