Research map / Image generation / Fast diffusion sampling
Rethinking Flow and Diffusion Bridge Models for Speech Enhancement
Dahan Wang, Jun Gao, Tong Lei, Yuxiang Hu, Changbao Zhu, Kai Chen, Jing Lu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace NCSN++ backbone with time-embedding-assisted TF-GridNet (ablation) | SBVE bridge model · DNS3 test set | PESQ: 2.802 -> 3.068 | Improved |
| use improved loss (SI-SNR plus power-compressed spectrum) instead of original data prediction loss (ablation) | NCSN++ SBVE bridge model · DNS3 test set | PESQ: 2.706 -> 2.802 | Improved |
| use improved loss instead of original loss (ablation) | NCSN++ SBVE bridge model · DNS3 test set | SI-SNR (dB): 14.158 -> 13.481 | Hurt |
| add CRP fine-tuning (ablation) | TF-GridNet SBVE bridge model · DNS3 test set | PESQ: 3.068 -> 3.213 | Improved |
| add CRP fine-tuning (ablation) | TF-GridNet SBVE bridge model · DNS3 test set | SI-SNR (dB): 16.646 -> 16.424 | Hurt |
| use OUVE path instead of SBVE (ablation) | TF-GridNet bridge model · DNS3 test set | PESQ: 3.068 -> 2.129 | Hurt |
| use BBED path instead of SBVE (ablation) | TF-GridNet bridge model · DNS3 test set | PESQ: 3.068 -> 2.800 | Hurt |
| use OT-CFM path instead of SBVE (ablation) | TF-GridNet bridge model · DNS3 test set | PESQ: 3.068 -> 2.834 | Hurt |
| use SB-CFM path instead of SBVE (ablation) | TF-GridNet bridge model · DNS3 test set | PESQ: 3.068 -> 3.102 | No clear effect |
| increase sampling steps from 1 to 5 (ablation) | proposed bridge model · DNS3 test set | PESQ: 3.185 -> 3.213 | Improved |
| one-step sampling instead of multi-step (ablation) | proposed bridge model · DNS3 test set | PESQ: 3.185 vs TF-GridNet 3.187 | No clear effect |
| use proposed bridge model instead of SBVE with NCSN++ | speech enhancement · DNS3 test set | PESQ: 2.592 -> 3.213 | Improved |
Closest papers by meaning
- Predictive–Generative Drift Decomposition for Speech Enhancement and Separation (NeurIPS 2026 · similarity 0.89)
- Shallow Diffusion for Fast Speech Enhancement (Student Abstract) (AAAI 2024 · similarity 0.89)
- Complex-Cycle-Consistent Diffusion Model for Monaural Speech Enhancement (AAAI 2025 · similarity 0.88)
- Denoising Diffusion Bridge Models (ICLR 2024 · similarity 0.87)
- DiffAR: Denoising Diffusion Autoregressive Model for Raw Speech Waveform Generation (ICLR 2024 · similarity 0.86)
- Generative Pre-training for Speech with Flow Matching (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
