Timbre-separated transcription on URMP 3mix
58.6F-score (Frame-level Separation)Note-Level Contrastive Clustering
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Note-Level Contrastive ClusteringTraining Data=MusicNet, Coupling Strategy=Rescale, Loss=InfoNCE [8]2025.09 | 58.6 | 59.8 | 76.9 | |
| Note-Level Contrastive ClusteringTraining Data=MusicNet, Loss=InfoNCE [8]2025.09 | 58.5 | 60 | 77.1 | |
| Note-Level Contrastive ClusteringTraining Data=MusicNet, Embedding Dimension=16, Loss=InfoNCE [8]2025.09 | 58 | 59.5 | 76.5 | |
| Note-Level Contrastive ClusteringTraining Data=MusicNet, Coupling Strategy=Shared residual block, Loss=InfoNCE [8]2025.09 | 57.2 | 57.5 | 74.1 | |
| TanakaBaseline=Tanaka et al. (2020), Loss=InfoNCE [8]2025.09 | 56.5 | 56.5 | 72.8 | |
| Note-Level Contrastive ClusteringTraining Data=MusicNet, Loss=L_affinity [6]2025.09 | 53.9 | 57.1 | 73.5 | |
| Note-Level Contrastive ClusteringTraining Data=Synthetic, Loss=InfoNCE [8]2025.09 | 40.9 | 45.1 | 58.1 |