Speaker recognition on DramaSR 532K 428K utterances (test)
87.79Accuracy (All)DramaSR-LRM
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| DramaSR-LRMtraining=SFT + RL, confidence sampling=true2026.07 | 87.79 | 85.22 | 90.37 | 87.62 | 88.92 | 85.7 | 76.65 | |
| DramaSR-LRMbackbone=Qwen3-8B, training=SFT + RL2026.07 | 86.93 | 84.94 | 88.91 | 87.45 | 87.77 | 84.12 | 76.95 | |
| Label Propagationmode=baseline2026.07 | 85.49 | 82.41 | 88.58 | 85.34 | 87.12 | 82.37 | 67.45 | |
| Qwen3-8Btraining=SFT, confidence sampling=true2026.07 | 82.7 | 65.88 | 89.21 | 81.54 | 83.98 | 82.19 | 71.14 | |
| Label-aware pyannoteassignment=Hungarian algorithm2026.07 | 79.82 | 81.44 | 78.22 | 82.61 | 80.42 | 72.82 | 62.48 | |
| Qwen3-8Btraining=SFT2026.07 | 75.22 | 76.21 | 84.6 | 72.63 | 76.67 | 76.97 | 68.61 | |
| Qwen3-8Bmode=direct use2026.07 | 27.4 | 31.66 | 23.17 | 27.45 | 27.26 | 27.62 | 28.65 | |
| Facial-aware Guessneighborhood assumption=true2026.07 | 22.54 | 23.3 | 21.75 | 23.21 | 22.43 | 21.53 | 20.78 |