JamendoMIDI

Multi-instrument MIDI transcriptions of the vocal-removed Jamendo tracks of the ICME 2026 Academic Text-to-Music (ATTM) Grand Challenge, introduced with MIDI-REPA as its symbolic training targets and produced with AMTFlow.

How it was made

We run our transcription pipeline (AMTFlow, --config paper) on ATTM's 55,701 vocal-removed Jamendo tracks and discard 4,624 tracks with unusable outputs, yielding JamendoMIDI: 51,077 audio tracks paired with instrument-stem MIDI annotations.

MuScriptor produces unstable predictions when conditioned on either no instrument tags or all instrument tags, so the conditioning set comes from multi-instrument recognition: the union of MATPAC++ and Music Flamingo, mapped to seven instrument families (piano, guitar, bass, drums, strings, winds and synthesizer). MuScriptor then transcribes each track with one instrument per detected family:

family	MIDI track (MuScriptor instrument)	GM program
piano	acoustic piano	0
guitar	clean electric guitar	26
bass	electric bass	33
drums	drums	drum channel
strings	violin	40
winds	brass section	61
synthesizer	synth pad	88

Splits

For training MIDI-REPA, we exclude the 682 Song Describer Dataset (SDD) tracks and 930 ATTM reference tracks that JamendoMIDI contains (13 tracks are in both), and reserve 444 tracks for loss monitoring, leaving 49,034 tracks (~3,161 hours).

split	tracks	use in MIDI-REPA
train	49,034	training
validation	444	loss monitoring
test	1,599	excluded from training: SDD and ATTM reference tracks

Fields

field	type	meaning
track_id	string	Jamendo track ID
split	string	train, validation or test
test_set	string	for test: attm, song_describer, or attm+song_describer
instruments	list of strings	the MuScriptor instruments the track was conditioned on
families	list of strings	the same, as instrument families
duration_s	float	track duration in seconds
midi	bytes	the Standard MIDI File

import io
import pretty_midi
from datasets import load_dataset

ds = load_dataset("vaclis/JamendoMIDI", split="validation")
row = ds[0]
midi = pretty_midi.PrettyMIDI(io.BytesIO(row["midi"]))
print(row["track_id"], row["families"], [(i.name, len(i.notes)) for i in midi.instruments])

Audio

Audio is not included. track_id is the Jamendo track ID, so the recordings can be obtained from MTG-Jamendo or Jamendo. The transcriptions were made from instrumental (vocal-removed) versions of these tracks. If you start from the original recordings, feel free to remove the vocals with any vocal-separation model (for example BS-RoFormer or Demucs) before pairing them with the MIDI.

Limitations

• These are automatic transcriptions, not human annotations, and contain errors.
• Each family is written as one instrument: every guitar is a clean electric guitar, every wind instrument a brass section.
• There are no vocal parts, because the source audio is vocal-removed.

License

The transcriptions are released under CC BY-NC-SA 4.0. The underlying recordings keep their own Jamendo Creative Commons licenses.

Citation

JamendoMIDI was introduced with MIDI-REPA. If you use it, please cite:

@misc{yu2026midirepa,
      title={MIDI-REPA: Improving Text-to-Music Generation with Symbolic Representation Alignment},
      author={Song-Ze Yu and Fang-Duo Tsai and Yi-Hsuan Yang},
      year={2026},
      eprint={2610.20967},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2610.20967},
}