Score Rearrangement

We propose a notation-level rearrangement method that changes the difficulty level of piano scores. Score conversion directly on the notation domain enables us to process musical information in the scores comprehensively.

Key Findings

• We can directly process musical scores on the notation domain using sequence models.
• We can train a single model that convert scores between multiple difficulty levels.
• Bar-major score token (ST+) performs better than staff-major score token (ST).

Overall Architecture

We tokenize score at notation level into sequence and translate them into another difficulty level, and then detokenize the result back into score. We use level tokens (at the beginning of source sequence) for difficulty conditioning. By changing these tokens, users can control the playing difficulty of scores!

Samples

Source score

• Intermediate (Lv.3)

Generated scores

• Beginner (Lv.1)

• Elementary (Lv.2)

• Advanced (Lv.4)

… and more samples to come!