Paper Detail
Junyan Jiang, Ruibin Yuan, Jiahao Pan, Wei Xue, Yike Guo, Gus Xia, Yann LeCun
Transcribing music into a human-readable score requires a coherent understanding of rhythm, harmony, melody, and form. Two obstacles limit this goal: annotated recordings are scarce, and accurate local predictions can still produce inconsistent musical sequences. We present SheetSage2, a unified music transcription framework that combines synthetic data, task-specific structured decoding, and autoregressive distillation. Automatically annotated MIDI, rendered into audio, provides scalable supervision across music understanding tasks. Task-specific structured decoders integrate complementary musical cues and their temporal dependencies to produce musically coherent scores. Autoregressive distillation further retains transcription accuracy without task-specific dynamic programming at inference. Across eight benchmark collections, a single SheetSage2-AR model exceeds the listed prior systems on 12 of 15 benchmark--metric pairs in our evaluation, substantially improving over SheetSage1 and surpassing task-specific models on several benchmarks. Model weights and inference code are publicly available.
No structured notes yet. Add `summary_sections`, `why_relevant`, `claim_impact`, or `next_action` in `papers.jsonl` to enrich this view.
No ranking explanation is available yet.
No tags.
@misc{jiang2026sheetsage2,
title = {SheetSage2: Coherent Lead-Sheet Transcription with Synthetic Supervision},
author = {Junyan Jiang and Ruibin Yuan and Jiahao Pan and Wei Xue and Yike Guo and Gus Xia and Yann LeCun},
year = {2026},
abstract = {Transcribing music into a human-readable score requires a coherent understanding of rhythm, harmony, melody, and form. Two obstacles limit this goal: annotated recordings are scarce, and accurate local predictions can still produce inconsistent musical sequences. We present SheetSage2, a unified music transcription framework that combines synthetic data, task-specific structured decoding, and autoregressive distillation. Automatically annotated MIDI, rendered into audio, provides scalable superv},
url = {https://huggingface.co/papers/2610.05336},
keywords = {huggingface daily},
eprint = {2610.05336},
archiveprefix = {arXiv},
}
{}