TuttiCorpus

A large-scale paired dataset of ABC notation scores and audio (WAV) for audio-to-score transcription and audio-score alignment research.

• 363,666 pieces, each with one full-length ABC score (.abc) and one full-length rendered audio (.wav), sharing the same anonymous ID
• ~13,000 hours of audio in total
• Audio: WAV, 48 kHz, stereo, 16-bit PCM
• Official piece-level train/valid/test splits (327,249 / 18,180 / 18,181) matching the pretraining setup of the TUTTI model
• Per-piece segmentation plans (measure-level, with time boundaries) plus ready-to-use scripts to reproduce the chunk-level training data

Two ways to use this dataset

The dataset is ~8.3 TB, so it ships in two forms.

A small streaming sample (300 pieces) in WebDataset format, for browsing in the dataset viewer and for prototyping without downloading anything:

from datasets import load_dataset

ds = load_dataset("pzzzzz/TuttiCorpus", "sample", streaming=True)
for item in ds["train"]:
    score = item["abc.txt"]        # str, the ABC notation
    audio = item["wav"]["array"]   # decoded waveform
    bpm   = item["json"]["bpm"]
    break

Decoding the audio needs torchcodec (pip install torchcodec). To read the scores only and skip audio decoding entirely, use ds["train"].decode(False) and treat wav as raw bytes.

The full corpus as 1,455 self-contained tar shards under shards/. Each shard holds 250 pieces and can be used on its own, so you do not have to download all 8.3 TB to start:

hf download pzzzzz/TuttiCorpus --repo-type dataset \
    --include "shards/shard-00000.tar" --local-dir .
tar -xf shards/shard-00000.tar          # -> abc/ and wav/

Repository layout

sample/                 # 300-piece WebDataset sample (the `sample` config above)
    train/train-*.tar   #   {id}.abc.txt + {id}.json + {id}.wav per sample
    valid/valid-*.tar
    test/test-*.tar
    sample_manifest.jsonl
shards/shard-00000.tar ... shard-01454.tar
                        # full corpus; each shard is self-contained and holds
                        # 250 pieces as abc/{id}.abc + wav/{id}.wav
metadata.jsonl          # one line per piece: {"id", "abc", "wav", "bpm"}
splits/
    train.jsonl / valid.jsonl / test.jsonl   # piece-level ID lists
    segmentation_plans.jsonl                 # per-piece segment boundaries
    tools/segment_abc.py, tools/segment_wav.py
    reference/                               # original pipeline scripts (archive)
    README.md                                # reproduction guide
preview/                # 30 loose sample pairs for quick browsing

Note that the shards are ordered by piece ID, not by split, so every shard contains a mix of train, valid and test pieces. Use the ID lists in splits/ to filter. This is also why the full corpus is not exposed as a loadable config: WebDataset derives a sample key by stripping the last extension, so abc/{id}.abc and wav/{id}.wav end up under two different keys and cannot be paired automatically.

Reproducing the chunk-level training data

The pretraining setup uses measure-aligned chunks of at most 14.5 seconds rather than full pieces. splits/segmentation_plans.jsonl records the exact measure ranges and time boundaries for every piece, and the two scripts in splits/tools/ reproduce them:

pip install abctoolkit soundfile tqdm
tar -xf shards/shard-00000.tar          # -> abc/ and wav/

python splits/tools/segment_abc.py --abc-dir ./abc \
    --plans splits/segmentation_plans.jsonl --out ./abc_chunks --workers 32
python splits/tools/segment_wav.py --wav-dir ./wav \
    --plans splits/segmentation_plans.jsonl --out ./wav_chunks --workers 32

The resulting abc_chunks/<id>/<id>#a-b.abc and wav_chunks/<id>/<id>#a-b.wav correspond one to one, with naming identical to the training pipeline. See splits/README.md for the format details and for how the plans were generated.

Notes

• All piece IDs are anonymized; original file and directory names are not included.
• bpm in metadata.jsonl is the notated tempo parsed from the score.
• The three official splits sum to 363,610 pieces. The remaining 56 pieces are present in the shards but were not assigned to any split, so filtering by the ID lists in splits/ will skip them.

License

cc-by-nc-4.0

Citation

Our paper has been accepted by ISMIR 2026!

The official camera-ready paper and arXiv preprint will be available soon. For now, if you use TuttiCorpus, our dataset, or code in your research, please cite us using the following BibTeX entry:

@inproceedings{hu2026tutti,
  title     = {TUTTI: Toward Generalizable Audio-to-Score Transcription via Fully Synthesized Data},
  author    = {Hu, Jianhuai and Wang, Yashan and Wu, Shangda and Guo, Zhancheng and Liang, Shijie and Meng, Wuna and Yang, Chuanqi and Li, Xiaobing and Yu, Feng and Sun, Maosong},
  booktitle = {Proceedings of the International Society for Music Information Retrieval Conference (ISMIR)},
  year      = {2026},
  note      = {To appear},
  url       = {https://github.com/a-musiclover/TUTTI}
}