2026:Music Structure Analysis Results
From MIREX Wiki
Submissions
| Submission | Title | Authors | |
|---|---|---|---|
| Baseline: SongFormer | SongFormer: Scaling Music Structure Analysis with Heterogeneous Supervision | Link | |
| SCALE | The SCALE System for MIREX 2026 Music Structure Analysis | Hongbing Li, Yaolong Ju | |
| songmamba | SongMamba: Efficient Music Structure Analysis via Bidirectional State-Space Modeling | Yuan Wang, Zhiping Qiu, Shenyang Xu, Yan Han, Yuanfeng Jiang, Xiaobing Li, Wei Li |
Test Set
- Harmonix Set (test split): 200 songs of the relabeled Harmonix Set (m-a-p/harmonixset_bigvgan, audio re-synthesized with BigVGAN), labels mapped to 7 classes (intro, verse, chorus, bridge, inst, outro, silence). These are the songs and audio of SongFormBench-HarmonixSet (identical files; annotations identical except for corrected endings of 2 songs).
For each metric, the best result is shown in bold.
Harmonix Set (test, 200 songs)
| Group | ACC | HR.5F | HR3F |
|---|---|---|---|
| SCALE | 0.799 | 0.728 | 0.828 |
| songmamba | 0.776 | 0.739 | 0.840 |
| Baseline: SongFormer | 0.802 | 0.704 | 0.798 |
| Baseline: MusicFM (MIREX 2025, reported) | 0.705 | 0.644 | 0.710 |
Notes
- Baseline SongFormer was also trained on SongFormDB, which may overlap with some popular song datasets to a small extent.
- SCALE was trained on SheetSage-MSA, which may overlap with some popular song datasets to a small extent.
