Agogic: Performance-Timed Music Tokens for LLM-Native Text-to-Symbolic-Music Generation
2026-08-04 • Sound
SoundComputation and Language
AI summaryⓘ
The authors studied how different ways of turning music into tokens (tokenization) affect text-to-music models, keeping all other factors the same. They found that the choice of representation matters more than model size for how well the generated music matches real music distributions. A new tokenization method they introduce, called PMT, greatly improves distribution fidelity even with much smaller models. They also show that improving certain qualities like instrument accuracy in generated music can be done without hurting overall distributional quality. Additionally, they provide datasets and tools to help others measure the impact of tokenization choices in music generation.
tokenizationtext-to-music modelsrepresentationFrechet Music Distance (FMD)PMT (Performance-Resolution Tokenization)model size scalingdistributional fidelitycaption adherenceinstrument-F1music generation datasets
Authors
Junhao Chen, Mingjin Chen, Jingjia Mao, Lin Chen, Saining Zhang, Minglin Chen, Ruocheng Wu, Liaoyuan Fan, Wenyi Li, Mingju Gao, Henghaofan Zhang, Zhihao Li, Hao Zhao, Yufei Wang, Ruqi Huang
Abstract
Text-to-music language models begin with a choice usually made by default: how to tokenize music. Normally entangled with backbone, data, and recipe, its effect has never been measured in isolation. We fix pretrained Qwen3.5 (0.8B-27B), data, budget, and decoding, and swap only the representation across seven tokenizations, anchoring texture metrics to each representation's model-free ceiling. The ordering is clean and surprising: representation, not model size, is the binding variable for distributional fidelity. Scaling the backbone 34x barely moves Frechet Music Distance (FMD), whereas switching representation halves it. PMT, a performance-resolution stream we release (10 ms timing, per-note velocity, multi-track texture; 609 symbols), reaches FMD 159 at 0.8B against 272-286 for beat grids (1.7-1.8x lower, up to 2.8x elsewhere; non-overlapping bootstrap CIs), so a 0.8B performance-resolution model beats a 27B beat grid. It reappears on a 26M from-scratch backbone and a second performance-resolution tokenizer: a property of the class, not one lucky vocabulary. Nor is it a finer-lattice artifact: snapping PMT's onsets to the beat grids' resolution still leaves it 67-129 FMD ahead of both (n=500). The effect is distributional; whether it is audible is a separate question, left open by our probe, with a human study pre-registered. Native caption adherence is weak but separable: a lightweight decode-time constraint doubles instrument-F1 (.28 to .60) and Correct-Key (.16 to .35) at no distributional cost. We release the harness, 25+ checkpoints, two corpora (86.6k aligned across caption/MIDI/ABC/audio; 6.25M captioned, the largest for music), and an imprinting diagnostic: published text-to-MIDI systems reproduce their training distribution near-invariant to the caption (72% vs. 71% chord-time on disjoint domains). The field's next representation claim can now be measured, not asserted.