Speech Enhancement on DNS Challenge Real-world recordings 2020
4.1SIGSGMSE++
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| SGMSE++Training dataset=VB-DMD2025.07 | 4.1 | 3.65 | 4.02 | 3.66 | 2.53 | — | — | |
| TeacherSampling steps=30, Training dataset=VB-DMD2025.07 | 4.08 | 3.62 | 3.89 | 3.57 | 2.61 | — | — | |
| ROSE-CDSampling steps=1, Training dataset=VB-DMD, RCD=true, Loss=Joint PESQ and SI-SDR2025.07 | 4.01 | 3.53 | 3.77 | 3.42 | 2.51 | — | — | |
| Ours (RCD + SI-SDR loss)Sampling steps=1, Training dataset=VB-DMD, RCD=true, Loss=SI-SDR2025.07 | 3.95 | 3.52 | 3.52 | 3.27 | 2.43 | — | — | |
| Ours (+ RCD)Sampling steps=1, Training dataset=VB-DMD, RCD=true2025.07 | 3.88 | 3.53 | 3.43 | 3.19 | 2.38 | — | — | |
| Ours (w/o RCD)Sampling steps=1, Training dataset=VB-DMD, RCD=false2025.07 | 3.67 | 3.36 | 3.32 | 3.03 | 2.28 | — | — | |
| FlowSE-GRPORL=GRPO2026.01 | 3.604 | — | 4.161 | 3.356 | — | — | — | |
| Ours (RCD + PESQ loss)Sampling steps=1, Training dataset=VB-DMD, RCD=true, Loss=PESQ2025.07 | 3.5 | 3.13 | 3.74 | 3.02 | 1.81 | — | — | |
| FlowSE-GRPORL=None2026.01 | 3.397 | — | 4.035 | 3.115 | — | — | — | |
| MetricGAN+Training dataset=VB-DMD2025.07 | 2.88 | 3.26 | 3.39 | 2.45 | 1.52 | — | — | |
| Conv-TasNetTraining dataset=VB-DMD2025.07 | 2.87 | 3.07 | 3.59 | 2.52 | 2.07 | — | — |