malerlab/bpsd-unirqvae3-unidac4-ytsv

BPSD score-image, audio and notation tokens (U-MusT)

Tokenized Beethoven Piano Sonata Dataset v2 for U-MusT — the test-only split, and the only corpus in the collection carrying all four modalities: score-image tokens, audio tokens, and LMX notation.

Because it is held out for evaluation, the image tokens here are not shift-augmented: they have shape (1, 1, H, W, 4), a single tokenization. The audio tokens retain the 9-variant stack.

BPSD ships no system-level image alignment, so the authors hand-annotated the test subset used in the paper. The audio derives from the four public-domain performances included in the Zenodo record. No audio is redistributed.

Token files contain shift variants, not a single tokenization

This is the most important thing to know before loading these files. A one-pixel offset in a score image, or a one-sample offset in audio, changes the discrete token assignment almost entirely. Training on a single tokenization therefore makes a model brittle to offsets it will certainly encounter at inference. Every token file here instead stores a stack of discretization-shift variants on its leading axes.

Image tokens (.pt, torch.int16):

Shape	Leading axes	Where it appears
(8, 4, H, W, 4)	8 horizontal x 4 vertical one-pixel shifts	scanned and photographed sources (yolo_shifted)
(6, 8, 4, H, W, 4)	6 key transpositions x 8 horizontal x 4 vertical	synthetically rendered sources (yolo_shifted_augmented)
(1, 1, H, W, 4)	none — a single tokenization	test-only splits (yolo)

H and W are the token grid of one musical system at 16x compression, and the trailing 4 is the RQ-VAE's four codebooks.

Audio tokens (.dac, Descript Audio Codec files): codes has shape (9, 4, T) — 9 temporal shift variants, 4 codebooks, T frames at 44.1 kHz mono. The 9 variants are sample offsets from -20 to +20 in steps of 5. Directories named mono_augmented carry the variant stack; the rare mono directories carry a single tokenization.

Index the leading axes to select a variant; index 0 is the unshifted tokenization in every case.

Full details of the augmentation scheme are in Appendix B of the supplementary material, published with the article at doi.org/10.1109/TASLPRO.2025.3648794.

This repository contains BOTH image-codec generations

Despite the unirqvae3 in its name, this repository ships image tokens from both codec generations, because BPSD was distributed with both and they are kept as-is:

Subtree	Codec	Files
image_tokens/<piece>/unirqvae3_f16_c1024_k4/yolo/	unirqvae3	438
image_tokens/images/unirqvae_f16_c1024_k4/yolo/	unirqvae	471

This is deliberate and useful: the unirqvae subtree is the only published image-token set of that generation, and unirqvae is what the strings Image-to-Audio and the Audio-to-Image checkpoints in malerlab/u-must expect. Every other published image-token dataset is unirqvae3 only. Pick the subtree that matches your checkpoint.

Both were verified empirically rather than trusted from their directory names, by decoding sample files with each codec and checking which produced legible notation.

Codec compatibility — read before pairing with a checkpoint

Image tokens are not interchangeable between codec generations. The unirqvae and unirqvae3 codebooks differ, so the same token index means different things. Pairing a checkpoint with tokens from the other generation produces meaningless output, silently.

Released run	Direction	Image codec it expects
run-20250225_062905-9n1554as	Image-to-Audio, piano	unirqvae3
run-20250130_150202-x9znhap2	Image-to-Audio, multi-instrument	unirqvae
run-20250128_025927-ks0ibl4v	Audio-to-Image, multi-instrument	unirqvae

Every published image-token dataset in this collection is unirqvae3, so only the piano Image-to-Audio run pairs with them directly. For the strings and Audio-to-Image runs, image tokens must be re-baked with unirqvae-ytsv using scripts/bake_image_tokens.py from the code repository. This matters most for the Audio-to-Image direction, where image tokens are the output target — evaluating it against unirqvae3 tokens measures nothing.

infer.py reads the tokenizer from each run's own files/config.yaml, so inference on a downloaded checkpoint selects the right codec automatically. The mismatch only arises when pairing a checkpoint with a token dataset by hand.

Licensing

Released under CC BY-NC-SA 4.0.

Note that this is more restrictive than the original corpus, which is CC BY 3.0. The original imposes no share-alike or non-commercial condition; those were added here because this tokenization is distributed alongside corpora that do impose them, and because the U-MusT models were trained on that combined mixture. If you only need this corpus, obtain it from its original source under its own, more permissive terms.

The full per-corpus license table, and the conflicts between them, are documented in the code repository. This is not legal advice — verify the current terms with the corpus publisher before redistributing or using commercially.

Citation

Please cite the U-MusT paper, and also the original corpus as required by its license.

@article{jung2026umust,
  title   = {U-MusT: A Unified Framework for Cross-Modal Translation of Score Images, Symbolic Music, and Performance Audio},
  author  = {Jung, Jongmin and Kim, Dongmin and Lee, Sihun and Cho, Seola and Soh, Hyungjoon and Bukey, Irmak and Donahue, Chris and Jeong, Dasaem},
  journal = {IEEE Transactions on Audio, Speech and Language Processing},
  volume  = {34},
  pages   = {1876--1891},
  year    = {2026},
  doi     = {10.1109/TASLPRO.2025.3648794}
}