System-level human MOS correlation on MusicEval (5-fold CV)
0.846Kendall's TauA3a (+LoRA)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| A3a (+LoRA)Backbone=MuQ, Tuning=LoRA2026.03 | 0.846 | 0.96 | 0.969 | — | |
| A1 (Frozen+MSE)Backbone=Frozen MuQ, Loss=MSE2026.03 | 0.839 | 0.957 | 0.96 | — | |
| A2 (Frozen+Ord.)Backbone=Frozen MuQ, Loss=Ordinal2026.03 | 0.839 | 0.953 | 0.963 | — | |
| A3b (+Contr.)Backbone=MuQ, Loss=Contrastive2026.03 | 0.838 | 0.955 | 0.967 | — | |
| A3c (+Unc.Wt.)Backbone=MuQ, Loss=Uncertainty Weighting2026.03 | 0.836 | 0.955 | 0.968 | — | |
| A4 (MERT-95M)Backbone=MERT-95M2026.03 | 0.822 | 0.946 | 0.948 | — | |
| FAD (VGGish)Backbone=VGGish2026.03 | 0.14 | — | — | 0.295 | |
| Audiobox Aes.2026.03 | — | 0.2 | — | — |