330 pieces, each with aligned audio + reference MIDI (in `musicnet_midis.tar.gz`) + per-note CSV labels. CSV-to-MusicXML conversion is mechanical.
License: **CC-BY 4.0**
330 pieces is comfortably above the 50-100 floor. Reference MIDIs are directly downloadable (2.6 MB tarball) and come from Classical Archives, Suzumidi, and HarfeSoft. Limitation: CC-BY (attribution required in any output render).
2026-08-01.
1,276 piano performances, ~200 h, captured on Yamaha Disklaviers during the International Piano-e-Competition. Audio (44.1-48 kHz 16-bit PCM stereo) + MIDI with ~3 ms alignment. Train/val/test split built in.
License: **CC BY-NC-SA 4.0** (verified on the Magenta page's "License" section)
Gold-standard for piano transcription. The MIDI is exactly what the Disklavier played (so it's true performance MIDI, not synthesized). The built-in splits keep the same composition out of multiple subsets. Limitation: **piano only** — not multi-instrument. Use as the piano-stratum of a multi-instrument eval, not as a multi-instrument eval by itself. Also **CC BY-NC-SA 4.0** — non-commercial share-alike. If the pipeline ships a commercial product, this dataset cannot be used as a holdout benchmark. (NC clause flagged for orchestrator.)
2026-08-01 — Magenta page reachable, SHA256 checksums published.
2100 multi-track pieces, each with mixture + stems + MIDI per stem. Each MIDI is the input to the synthesizer, so it is exact ground truth.
License: **CC-BY 4.0**
2100 pieces is 20-40× the 50-100 floor. Multi-instrument (every mix has at least piano/guitar/drums/bass; median track has 4-8 sources). Limitation: known duplicate-MIDI issue — the Zenodo "redux" version moves duplicates to an `omitted/` directory; the `omitted/` tracks must NOT be used for transcription evaluation per the dataset's own guidance.
2026-08-01.
240,000 synthesized chamber-ensemble pieces (~1,411 h of audio), performed by 13 instruments (violin, viola, cello, double bass, flute, oboe, clarinet, bassoon, saxophone, trumpet, french horn, trombone, tuba) across 4 ensemble types (string, brass, woodwind, random). Each example: mixture audio + per-source audio + aligned MIDI + multi-f0 + per-note expression (vibrato, loudness, brightness). "Tiny" subset is 24k train / 8k val / 8k test.
License: **CC-BY 4.0** (verified on the Magenta page)
This is the largest multi-instrument AMT benchmark in existence. MIDI ground truth is exact (synthesizer input). "Tiny" subset is reasonable for quarterly eval. Limitation: 2.9 TB full / 4 TB uncompressed — you almost certainly want the tiny version. Synthesized, so it does not exercise real performance noise. CC-BY is permissive.
2026-08-01 — Magenta page and the GitHub data_download shell scripts both confirm CC-BY 4.0 and 2.9 TB / 569 GB main sizes.
44 multi-instrument pieces with MIDI scores + per-instrument audio + assembled mix + video.
License: **CC-BY 4.0**
Below the 50-100 floor (44 < 50), but the pieces are real human performances (not synthesized) and include score-engraved PDFs as well as MIDI. Suitable as a "real performance" supplement to a larger synthesized set. Limitation: 44 pieces is the entire dataset, and pieces are reused in the literature; you should be aware of evaluation contamination if the pipeline trains on URMP.
2026-08-01.
~30 GB of synthesized piano audio (Disklavier + synth), ~30 h, with ground-truth MIDI for every audio file. Seven subsets (AkPnBcht, AkPnBsdf, etc.) plus the augmented A-MAPS version.
License: **CC BY-NC-SA 2.0 (France)** — a non-commercial French-jurisdiction variant. The same dataset is also covered by the A-MAPS augmentation paper (CC-BY 4.0) but that only covers the A-MAPS metadata, not the audio itself.
Long-standing AMT benchmark; many papers report results on MAPS so this gives a "comparison to literature" anchor. Limitation: **piano only** AND **CC BY-NC-SA 2.0 (France)** — the French NC-SA license is more restrictive than the international CC BY-NC-SA 4.0 (different jurisdiction, different terms). Flag for orchestrator if commercial shipping is on the roadmap.
2026-08-01 — Zenodo mirror verified, license page in the host confirms CC BY-NC-SA 2.0.
500 Bach chorales (4-part SATB) in TSV, MusicXML, Humdrum kern, and MuseScore formats. Includes notes, measures, chords, harmonic labels. The original Bach works are public domain; the encoding is the question.
License: **LICENCE CONFLICT — see Risks/Gaps.** The GitHub repo README says `CC0 1.0 Universal (CC0-1.0)`; the documentation page says `CC BY-NC-SA 4.0`. The DCMLab distant_listening_corpus (sibling repo) is CC BY-NC-SA 4.0, so the CC0 claim on the Bach chorales repo looks anomalous — possibly a copy-paste error from a generic README template, or possibly intentional for the chorales specifically. Need orchestrator decision before adopting.
Best-in-class symbolic ground truth: 500 SATB chorales with multiple annotation facets. The pipeline's music21 + Verovio stack is purpose-built for this kind of data. Limitation: **the audio is not provided** — these are symbolic scores only. You would need a separate synthesizer (or a YAMAHA Disklavier, in your dreams) to turn the MusicXML into audio. So this is a ground-truth source, not an audio+ground-truth source. Also the license discrepancy above.
2026-08-01 — both the dcmlab.github.io page and the GitHub README were fetched; the conflict is real.
Built-in corpus shipped with `music21` (already in the pipeline's `trusted_packages.txt` stack). Includes the complete Bach chorales, many Haydn/Beethoven string quartets, three books of Monteverdi madrigals, thousands of folk songs (Essen, ABC). All encoded in MusicXML / Humdrum / MuseScore.
License: **Mixed, per composition** — music21's "Authors, Acknowledgments" page states: *"the music (if not the encodings) in the corpus are either out of copyright in the United States and/or are licensed for non-commercial use... Some encodings included in the corpus may not be used for commercial uses or have other restrictions."* No blanket license — you must check the per-work README in the corpus tree.
Zero install overhead (it's already in the project) and the works themselves are mostly PD (Bach chorales pre-1929; Beethoven quartets; etc.). Limitation: per-work license check is required; some encodings are NC. Also no audio.
2026-08-01 — music21 docs site reachable, licensing section quoted.
176,581 unique MIDI files (LMD-full), 45,129 of which are matched to Million Song Dataset entries (LMD-matched). LMD-aligned is a 272 MB subset pre-aligned to MSD. Multi-instrument, multi-genre, no audio.
License: **CC-BY 4.0** (confirmed on Colin Raffel's project page and on the MusPy loader metadata: `_LICENSE = "Creative Commons Attribution 4.0 International License (CC-By 4.0)"`)
Massive scale — 176k files = essentially unlimited symbolic ground truth. The `lmd_aligned` subset is the practical entry point. Limitation: **no audio** — these are MIDI files only. To build an audio+ground-truth eval you'd need to synthesize (Slakh is exactly this — Slakh is the audio rendering of a subset of LMD). LMD-full contains corrupt files; the `_clean` subset is safer.
2026-08-01 — Colin Raffel project page and MusPy `lmd.py` source both confirm CC-BY 4.0.
GMD: 13.6 h / 1,150 MIDI files of human-played drum performances on a Roland TD-11 V-Drum, with synthesized audio. E-GMD: 444 h, 43 drum kits, on Roland TD-17.
License: **CC-BY 4.0** for both GMD and E-GMD (verified on both Magenta pages)
Best-in-class drum transcription ground truth. Use as the "drums" stratum of a multi-instrument eval (combine with Slakh stems for the harmonic content). Limitation: **drums only** — not a general multi-instrument source. The 22,000 drum measures in GMD give plenty of eval material; E-GMD's 444 h is overkill for a quarterly eval.
2026-08-01.
869,000+ scores, 94,000+ recordings, 27,000+ composers, 258,000+ works. The de-facto public-domain classical score library. Has MIDI for many works in the user-uploaded score collection.
License: **Mixed, per work.** Per-work license is shown on every score page. Generally: scores are public domain for composers who died before 1972 (Canada jurisdiction) / 1955 (EU) / 1975 (Taiwan); recordings have separate copyright and many are *not* PD. *For a non-PD score, IMSLP requires the copyright holder's permission before upload.* Site content is CC-BY 4.0 (re3data).
Vast and the works are PD where composers are pre-1929-dead, which is most of the standard classical repertoire. Limitation: license check is per-page; *recordings* on IMSLP are NOT generally PD; for AMT you really want the recording, not just the score. So IMSLP is best used as a score-side resource to pair with separately-sourced audio.
2026-08-01 — IMSLP main page and "Copyright Made Simple" page both fetched; ruleset confirmed.
108 multitrack choral pieces with aligned audio, F0, beats, and notes. 4-part (SATB) and 8-part arrangements. Synthesized via a singing-voice synthesizer, not real human singers.
License: **CC-BY 4.0** (per mirdata 1.0.0 quick reference, which fetches from the dataset's own Zenodo record)
Choral, not general "multi-instrument" — but a great match if the pipeline ever extends to vocal ensemble. Limitation: synthesized audio; not a true cross-genre test.
2026-08-01 — mirdata quick-reference fetched and cross-checked.
1,186,253 MIDI files, ~100,629 h, derived from audio via AMT (Lakh is one input). Largest MIDI collection to date.
License: Per the arXiv paper (Bradshaw et al., 2025) and the GitHub repo, Aria-MIDI itself is **research use; check repo LICENSE for redistribution rights**. The arXiv paper says over 1 million distinct MIDI files totaling ~100,000 hours of transcribed audio.
Scale is unmatched. Limitation: derived from AMT, so the MIDI is itself the output of a transcription model — using AMT-derived MIDI as AMT ground truth is somewhat circular; you'd be measuring transcription-vs-transcription rather than transcription-vs-human. License is less clear than CC-BY; need orchestrator sign-off.
2026-08-01 — arXiv abstract and GitHub repo both fetched.
None
License: —
None
None
None
License: —
None
None