notesum.ai
Published at October 23Exploring Tokenization Methods for Multitrack Sheet Music Generation
cs.CV
cs.AI
cs.LG
Released Date: October 23, 2024
Authors: Yashan Wang1, Shangda Wu1, Xingjian Du2, Maosong Sun3
Aff.: 1Central Conservatory of Music, China; 2University of Rochester, USA; 3Central Conservatory of Music, China; Tsinghua University, China

| Tokenization | Parameters | Sec/Epoch | Inference Speed | BPB | CLaMP 2 Score | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Bach | Haydn | Mozart | Bach | Haydn | Mozart | Bach | Haydn | Mozart | |||
| Byte patching | 65,872,896 | 963 | 597.1 | 630.4 | 623.8 | 0.2795 | 0.3682 | 0.3900 | 0.9767 | 0.9071 | 0.8068 |
| Line-stream patching | 65,872,896 | 1107 | 549.7 | 564.8 | 569.0 | 0.2772 | 0.3797 | 0.3958 | 0.9734 | 0.8916 | 0.8213 |
| Bar-stream patching | 65,872,896 | 1063 | 446.3 | 465.6 | 449.6 | 0.2539 | 0.3526 | 0.3879 | 0.9781 | 0.9228 | 0.8225 |
| Bar patching | 70,628,352 | 2848 | 226.1 | 210.9 | 204.3 | 0.2479 | 0.3515 | 0.3920 | 0.9813 | 0.9045 | 0.7531 |
| BPE | 84,074,496 | 4071 | 91.0 | 80.2 | 71.1 | 0.2591 | 0.3340 | 0.3542 | 0.9687 | 0.9050 | 0.7005 |