Research map / Image generation / Fast diffusion sampling
Shallow Diffusion for Fast Speech Enhancement (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SDFEN, a shallow diffusion speech enhancement method that starts the reverse process from an adaptively chosen intermediate step with a learned noisy predictor, generating speech in two steps.
Problem
Diffusion-based speech enhancement needs many reverse iterations, causing high computational cost and slow inference.
Approach
Initiate the reverse diffusion process at an intermediate time step using a learned noisy predictor that yields a prior from the noisy speech, switch to x0-space prediction training, and reduce sampling to two steps.
Key results
- SDFEN achieves PESQ 2.55 and CSIG 3.81 on VoiceBank-DEMAND with only 2 reverse steps, beating all diffusion baselines on PESQ and CSIG.
- SDFEN's STOI is 93.4, slightly below CDiffuSE (93.7) and DiffuSE (93.5), and CBAK 3.27 ties DR-DiffuSE.
- SGMSE with 50 steps underperforms methods using fewer steps, which the authors take as evidence that starting from very noisy steps does not help.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SDFEN's shallow diffusion and two-step inference instead of DR-DiffuSE's six-step inference | DiffWave speech enhancement · VoiceBank-DEMAND | reverse steps: 6 -> 2 | Helped |
| Use SDFEN instead of DR-DiffuSE | DiffWave speech enhancement · VoiceBank-DEMAND | PESQ: 2.50 -> 2.55 | Helped |
| Use SDFEN instead of CDiffuSE | DiffWave speech enhancement · VoiceBank-DEMAND | STOI: 93.7 -> 93.4 | Hurt |
| Use SDFEN instead of CDiffuSE | DiffWave speech enhancement · VoiceBank-DEMAND | CSIG: 3.77 -> 3.81 | Helped |
| Use SDFEN instead of CDiffuSE | DiffWave speech enhancement · VoiceBank-DEMAND | CBAK: 3.09 -> 3.27 | Helped |
| Use SDFEN instead of DR-DiffuSE | DiffWave speech enhancement · VoiceBank-DEMAND | CBAK: 3.27 -> 3.27 | No clear effect |
| Use SDFEN instead of the unprocessed input | Speech enhancement · VoiceBank-DEMAND | PESQ: 1.97 -> 2.55 | Helped |
Surprising findings
- SGMSE with 50 reverse steps yields lower scores than baselines using only 6 steps.
Limitations
- Evaluated only on VoiceBank-DEMAND
- T1 and t2 are hand-chosen hyperparameters tuned on a validation subset
- Does not beat all baselines on STOI
Experimental setup
Datasets: VoiceBank-DEMAND
Benchmarks: VoiceBank-DEMAND
Metrics: STOI PESQ CSIG CBAK number of reverse steps
Baselines: DiffuSE CDiffuSE SGMSE DR-DiffuSE
Builds on: DiffuSE DiffWave SGMSE
Closest papers by meaning
- Complex-Cycle-Consistent Diffusion Model for Monaural Speech Enhancement (AAAI 2025 · similarity 0.88)
- DiffAR: Denoising Diffusion Autoregressive Model for Raw Speech Waveform Generation (ICLR 2024 · similarity 0.88)
- Cauchy Diffusion: A Heavy-tailed Denoising Diffusion Probabilistic Model for Speech Synthesis (AAAI 2025 · similarity 0.87)
- Predictive–Generative Drift Decomposition for Speech Enhancement and Separation (NeurIPS 2026 · similarity 0.87)
- DDDM-VC: Decoupled Denoising Diffusion Models with Disentangled Representation and Prior Mixup for Verified Robust Voice Conversion (AAAI 2024 · similarity 0.87)
- SADA: Stability-guided Adaptive Diffusion Acceleration (ICML 2025 · 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.