2026:Symbolic Music Generation Results
Submissions[edit]
| Team | Extended Abstract | Methods |
|---|---|---|
| DAMN (Langchu Huang, Yang Xu) | Nested Transformer | |
| PhraseSplitGen (Di Lu, Xinran Zhang, Qingwen Zhou) | Transformer | |
| MuseCoco (BL-1) | [1] | Transformer |
| Anticipatory Music Transformer (BL-2) | [2] | Transformer |
Interactive Demo[edit]
Play-along piano-roll continuations (visual + audio) for every prompt and system are available on the interactive demo page.
Results[edit]
| Team | Subjective Evaluation | |||
|---|---|---|---|---|
| Coherecy ↑ | Structure ↑ | Creativity ↑ | Musicality ↑ | |
| DAMN | 3.68 ± 0.16a | 3.39 ± 0.18a | 3.18 ± 0.16a | 3.47 ± 0.14a |
| PhraseSplitGen | 2.95 ± 0.18b | 2.74 ± 0.15b | 2.89 ± 0.15a | 2.61 ± 0.14b |
| MuseCoco (BL-1) | 2.79 ± 0.17b | 2.76 ± 0.15b | 3.42 ± 0.12a | 2.68 ± 0.16b |
| Anticipatory Music Transformer (BL-2) | 3.47 ± 0.20a | 3.34 ± 0.16a | 3.34 ± 0.16a | 3.61 ± 0.17a |
Evaluation Results
Results are reported in the form of mean ± sems (sem refers to standard error of mean), where s is a letter. Different letters within a column indicate significant differences (p-value p < 0.05) based on a Wilcoxon signed-rank test with Holm-Bonferroni correction.
Baseline Models
For MuseCoco, we use the xlarge model with 1.2 billion learnable parameters. For Anticipatory Music Transformer, we use the Large model with 780M learnable parameters.
Subjective Evaluation Details
A double-blind online survey was conducted to test music quality. Each model was anonymised, and for each test prompt, a sample was cherry-picked from 8 generated candidates. A total of 8 prompts of varied styles (pop, classical, and jazzy) were tested, resulting in an 8-page survey. The page order and the sample order within each page were both randomised.
We collected a total of n=38 ratings for each model across 12 survey participants with diverse music backgrounds. The average completion time was 20 minutes.