Reproduce “Score Transformer: Generating Musical Score from Note-level Representation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Masahiro Suzuki

The Score Transformer, developed by Masahiro Suzuki at Yamaha Corporation, automatically generates comprehensive musical scores from note-level MIDI data. This system significantly improves upon existing commercial software and research baselines in accurately notating complex musical elements such as voice assignment, beaming patterns, and stem directions.

Key insights

  • Treating comprehensive musical score generation as a sequence-to-sequence translation problem enables transformer models to learn complex notation conventions from data.
  • A custom tokenization scheme specifically designed for musical notation allows neural networks to effectively process and generate detailed score elements.
  • Training the model on a 'score restoration' task, where original scores are down-converted to note-level data and then reconstructed, is an effective supervised learning approach.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.