Pianoroll-Event: A Novel Score Representation for Symbolic Music
Published in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2026), 2026
Overview of the Pianoroll-Event encoding pipeline: from pianoroll to compressed event sequence.
Abstract
We introduce an encoding scheme that merges pianoroll representations with event-based descriptions. Our approach defines four event types:
| Event Type | Description |
|---|---|
| Frame Events | Capture temporal structure of musical segments |
| Gap Events | Handle silence and rests between notes |
| Pattern Events | Encode repetitive musical patterns efficiently |
| Structure Events | Represent high-level musical form |
This scheme achieves 1.36x to 7.16x encoding efficiency compared to comparable discrete sequence approaches. Testing across multiple autoregressive models demonstrates consistent performance gains in both objective metrics and human evaluation.
Links
Recommended citation: Lekai Qian, Haoyu Gu, Dehan Li, Boyu Cao, Qi Liu. (2026). "Pianoroll-Event: A Novel Score Representation for Symbolic Music." ICASSP 2026.
Download Paper