Corrective Forcing: Unified Post-Training for Diffusions and Flows in Generative Speech EnhancementarXiv preprint:2609.24651
← Home

Paper Demo

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

arXiv preprint:2609.24651

Qing Yao[email protected]Lijian Gao[email protected]Qirong Mao†[email protected]

  1. Jiangsu University

† Corresponding author

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.

Key Words

Generative speech enhancementSchrödinger bridgeFlow matchingPost-training methodsExposure bias

Audio Samples

Listen to 10 samples from both datasets. Every sample compares the pretrained formulation with its CoF post-trained counterpart. Use the sampling-budget toggles to hear the same samples at 1 NFE and 4 NFE.

WSJ0+WHAM

Denoising5 samples

Speech enhancement under additive environmental noise.

Visualization
#1Sample name422c020d_c1_2103_snr=-1.7.wavSNR-1.7 dBSignal-to-noise ratio (SNR) measures speech strength relative to background noise. Higher is cleaner.
Sampling budget
Methods
0:00
-0:00
100
0–200

WSJ0+REVERB

Dereverberation5 samples

Speech enhancement under simulated room reverberation.

Visualization
#1Sample name052o020d_c1_698_t60=0.71.wavT600.71 sReverberation time (T60) is the time for sound to decay by 60 dB. Lower means less reverberation.
Sampling budget
Methods
0:00
-0:00
100
0–200

Results

Compare pretrained and CoF post-trained models on WSJ0+WHAM (denoising) and WSJ0+REVERB (dereverberation). Chart view benchmarks all five metrics at one sampling budget, and Curves view traces each metric over NFE (number of function evaluations).

Benchmark comparison

Dataset
View
Sampling
  • Unprocessed
  • Clean reference
  • SGMSE+@100
  • StoRM@100+1
  • SB-VE@4
  • CoF (SB-VE)@4
  • OT-CFM@4
  • CoF (OT-CFM)@4
1.011.552.092.623.161.32Unprocessed, NFE —: 1.32 2.14SGMSE+, NFE 100: 2.14 2.53StoRM, NFE 100+1: 2.53 2.41SB-VE, NFE 4: 2.41 2.83CoF (SB-VE), NFE 4: 2.83 2.56OT-CFM, NFE 4: 2.56 2.85CoF (OT-CFM), NFE 4: 2.85
PESQ Perceived quality against clean speech, scored −0.5 to 4.5. Higher is better.
0.490.580.670.760.850.55Unprocessed, NFE —: 0.55 0.71SGMSE+, NFE 100: 0.71 0.76StoRM, NFE 100+1: 0.76 0.78SB-VE, NFE 4: 0.78 0.80CoF (SB-VE), NFE 4: 0.80 0.78OT-CFM, NFE 4: 0.78 0.80CoF (OT-CFM), NFE 4: 0.80
ESTOI Intelligibility against clean speech, scored 0 to 1. Higher is better.
1.65.69.513.517.43.9Unprocessed, NFE —: 3.9 dB10.3SGMSE+, NFE 100: 10.3 dB13.4StoRM, NFE 100+1: 13.4 dB14.3SB-VE, NFE 4: 14.3 dB15.2CoF (SB-VE), NFE 4: 15.2 dB14.1OT-CFM, NFE 4: 14.1 dB15.1CoF (OT-CFM), NFE 4: 15.1 dB
SI-SDR (dB) Signal fidelity against clean speech in dB, where negative values are possible. Higher is better.
2.572.782.993.213.422.69Unprocessed, NFE —: 2.69 2.95SGMSE+, NFE 100: 2.95 3.12StoRM, NFE 100+1: 3.12 3.27SB-VE, NFE 4: 3.27 3.26CoF (SB-VE), NFE 4: 3.26 3.30OT-CFM, NFE 4: 3.30 3.25CoF (OT-CFM), NFE 4: 3.25
UTMOS Predicted quality without a reference, scored 1 to 5. Higher is better.
2.907.8012.6917.5822.4816.10Unprocessed, NFE —: 16.10 %19.68SGMSE+, NFE 100: 19.68 %15.40StoRM, NFE 100+1: 15.40 %13.73SB-VE, NFE 4: 13.73 %11.56CoF (SB-VE), NFE 4: 11.56 %13.57OT-CFM, NFE 4: 13.57 %12.06CoF (OT-CFM), NFE 4: 12.06 %5.70Clean reference, NFE : 5.70 %
WER (%) Word error rate against the reference transcript, 0% and up. Lower is better.

Citation

BibTeX
@article{yao2026cof,
  author  = {Yao, Qing and Gao, Lijian and Mao, Qirong},
  title   = {Corrective Forcing: Unified Post-Training for Diffusions and Flows in Generative Speech Enhancement},
  journal = {arXiv preprint arXiv:2609.24651},
  year    = {2026}
}