Reproduce “BEAT: Tokenizing and Generating Symbolic Music by Uniform Temporal Steps” 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

Lekai Qian, Haoyu Gu, Jingwei Zhao, Ziyu Wang

Researchers from South China University of Technology, National University of Singapore, Mohamed bin Zayed University of Artificial Intelligence, and New York University introduced BEAT, a novel tokenization framework for symbolic music that explicitly encodes musical time in uniform beat-length steps, enhancing generative Transformer models. The framework produces music with superior rhythmic consistency and structural coherence, outperforming existing event-based methods in objective and subjective evaluations, while also enabling real-time conditional generation.

Key insights

  • Explicitly encoding uniform temporal steps as an inductive bias significantly improves a Transformer's ability to learn and generate musically coherent structures, especially rhythmic regularity and long-range patterns.
  • BEAT's design, which provides explicit and localized beat boundaries, enhances computational efficiency through compact sequences and facilitates robust conditional and real-time generation by preventing information leakage across beats.
  • Approximating pitch position through relative indexing within a beat and summarizing velocity as a single mean value per pitch offers a near-lossless and compact representation that is approximately invariant to transposition and rhythm shifts, aiding learning and generalization.

Summary by alphaXiv.

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