BEAT: Rethinking Symbolic Music Encoding with Beat-Anchored Tones Representations

The perception of symbolic music is fundamentally based on the relative durations and interval relationships between the notes. However, existing representations face an inherent trade-off between preserving these musical priors and achieving encoding efficiency. In this paper, we propose BEAT (Beat-based Encoding with Anchored Tones), which breaks from the conventional note-event paradigm by adopting beat units as the basic primitive. BEAT segments the temporal axis into beat units, organizing music as sequences of these units. Within each beat, notes are sorted by pitch, with the first serving as an anchor point, while subsequent notes are encoded through chained relative intervals coupled with rhythmic patterns. This approach achieves efficient serialization while explicitly preserving musical priors.

Additionally, BEAT employs beat indexing for temporal alignment, enhancing both concurrent modeling capabilities and editability, enabling seamless adaptation to diverse downstream tasks. Extensive experiments demonstrate that BEAT achieves better encoding efficiency than existing methods while achieving state-of-the-art generation quality in both objective and subjective benchmarks. BEAT also consistently outperforms baselines across diverse tasks including multi-track generation, segment infilling, and real-time accompaniment.