Personalized Speech Enhancement on DNS Track 2: Speakerphone Blind 5 (test)
4.73SIG ScoreUniSE + GSPO
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| UniSE + GSPOPost-training=GSPO2026.06 | 4.73 | 4.76 | 4.61 | |
| GenSE + GSPOPost-training=GSPO2026.06 | 4.61 | 4.53 | 4.36 | |
| UniSE2026.06 | 4.47 | 4.41 | 4.15 | |
| UniFlowDDPM2026.06 | 4.09 | 3.88 | 3.56 | |
| LLaSE-G12026.06 | 4.08 | 3.84 | 3.55 | |
| UniFlowFM2026.06 | 4.06 | 3.89 | 3.54 | |
| NoisyReference Speaker Embedding=-, On-the-fly Speaker Embedding=-2026.01 | 4.05 | 2.16 | 2.5 | |
| Noisy2026.06 | 4.05 | 2.16 | 2.5 | |
| UniFlowMF2026.06 | 4.04 | 3.87 | 3.51 | |
| TEA-PSEVersion=3.02026.06 | 3.99 | 3.95 | 3.49 | |
| NAPSE2026.06 | 3.92 | 4.17 | 3.56 | |
| GenSE2026.06 | 3.92 | 3.07 | 2.96 | |
| E3NetReference Speaker Embedding=ECAPA, On-the-fly Speaker Embedding=-, Params (M)=6.62, MACs (G)=11.082026.01 | 3.72 | 3.17 | 2.87 | |
| pDCCRNReference Speaker Embedding=ECAPA, On-the-fly Speaker Embedding=-, Params (M)=4.50, MACs (G)=5.592026.01 | 3.65 | 3.09 | 2.81 | |
| pDeepFilterNet2 (oracle similarity)Reference Speaker Embedding=ECAPA, On-the-fly Speaker Embedding=ECAPA, scaling factor alpha=2, Params (M)=17.67, MACs (G)=1.752026.01 | 3.64 | 3.04 | 2.77 | |
| pDeepFilterNet2 (light similarity)Reference Speaker Embedding=ECAPA, On-the-fly Speaker Embedding=TinyECAPA, scaling factor alpha=6, Params (M)=2.38, MACs (G)=0.372026.01 | 3.63 | 3.19 | 2.83 | |
| pDeepFilterNet2 (oracle similarity w/o scaling)Reference Speaker Embedding=ECAPA, On-the-fly Speaker Embedding=ECAPA, scaling factor alpha=1, Params (M)=17.67, MACs (G)=1.752026.01 | 3.59 | 3.12 | 2.78 | |
| pDeepFilterNet2Reference Speaker Embedding=ECAPA, On-the-fly Speaker Embedding=-, Params (M)=2.23, MACs (G)=0.332026.01 | 3.58 | 3.17 | 2.78 | |
| pDeepFilterNet2 (light similarity w/o scaling)Reference Speaker Embedding=ECAPA, On-the-fly Speaker Embedding=TinyECAPA, scaling factor alpha=1, Params (M)=2.38, MACs (G)=0.372026.01 | 3.53 | 3.3 | 2.81 |