Guides · checked 1 Nov 2026
Open music datasets and what their licences allow
Most music datasets you can download are licensed for research only. Here is each one, what it holds, its licence and whether a commercial model may train on it, read from the dataset's own page on 1 Nov 2026.
| Dataset | Holds | Licence | Commercial training |
|---|---|---|---|
| MUSDB18 | 150 songs, about 10 h, mix plus four stems (vocals, drums, bass, other), encoded audio | Educational use only; no commercial use without the copyright holders' permission | No |
| MUSDB18-HQ | The same 150 songs as uncompressed WAV | Educational use only, as MUSDB18 | No |
| MoisesDB | 240 songs from 45 artists in 12 genres, mix plus stems per instrument | CC BY-NC-SA 4.0 | No |
| MedleyDB 1.0 and 2.0 | 196 multitracks (122 plus 74): mix, processed stems and raw audio, with annotations | CC BY-NC-SA 4.0, non-commercial research use only | No |
| Slakh2100 | 2,100 songs, 145 h, stems rendered from MIDI with sample libraries, MIDI aligned | CC BY 4.0 | Yes synthetic renders |
| Cambridge-MT (Mixing Secrets) | Raw multitrack sessions from the Mixing Secrets library | Educational purposes as agreed with the contributors; no formal licence | No ask each contributor |
| DSD100 | 100 songs, mix plus four stems, the SiSEC 2016 set | None stated; refers to the Mixing Secrets terms | Not stated |
| FMA | 106,574 full tracks, 343 days of audio | Audio under the licence each artist chose; metadata CC BY 4.0 | Per track |
| MTG-Jamendo | 55,000 full tracks with tags | Non-commercial research and academic use only; metadata CC BY-NC-SA 4.0 | No written permission from Jamendo |
| Lakh MIDI | 176,581 MIDI files collected from the web, no audio | CC BY 4.0 for the collection | Yes MIDI only; the songs themselves are not cleared |
| MAESTRO | 1,276 piano performances, 198.7 h, audio aligned with MIDI | CC BY-NC-SA 4.0 | No |
| GuitarSet | 360 guitar excerpts of about 30 s, about 3 h, with annotations | CC BY 4.0 | Yes |
| MusicNet | 330 freely licensed classical recordings with note labels | CC BY 4.0 | Yes |
| URMP | 44 classical pieces, one recording per instrument, MIDI score and video | None stated | Not stated |
| RWC Music Database | 315 pieces across popular, classical and jazz, plus instrument sounds | Research-only pledge | No |
| MusicCaps | 5,521 captions for 10 s YouTube clips, no audio | CC BY-SA 4.0 for the captions | Per track captions only; the audio is not covered |
| Song Describer | About 1,100 captions for 706 recordings | CC BY-SA 4.0 for the captions; audio per track, about seven in ten non-commercial | Per track |
| Million Song Dataset | Features and metadata for a million tracks, no audio | Echo Nest API terms; tags CC BY-NC-SA 2.0 | Not stated no audio |
Each name links to the page the licence was read from. Licences change and pages move. Check the source before you rely on a line, and tell us at mikas@ximaudio.com if one is out of date.
The pattern
Every set of real multitrack recordings of songs is non-commercial or has no licence at all: MUSDB18, MedleyDB, MoisesDB, Cambridge-MT, DSD100, URMP, RWC. The sets with a permissive licence are a different kind of thing. Slakh is rendered from MIDI with sample libraries, so the stems are synthetic. Lakh is MIDI with no recordings, and the songs behind the files were never cleared. GuitarSet and MusicNet are single instruments or classical pieces. MTG-Jamendo is non-commercial outright. FMA is full mixes under the licence each artist chose, and many of those forbid commercial use or derivative works. Song Describer calls its recordings permissively licensed; its own licence file shows about seven in ten are non-commercial.
A per-track set can be used commercially, but only the permissive tracks, and only if you keep the record of which ones. A non-commercial set lets you train a research model and publish the paper. It does not let you ship a product trained on it, and a buyer's diligence team reads it that way.
What is missing
There is no open set of real multitrack recordings that a commercial model may train on. Teams that need one either record it themselves, run a separation model over a licensed catalogue of mixes and train on the guesses, or license production data from the people who made it.
Where licensed multitrack data comes from
XIM Audio licenses the parts a song is made from, printed by the producer from the session it was made in. 4,223 songs, 110,662 lanes, about 9,153 hours, measured 4 Nov 2026. One 24-bit WAV per lane with an instrument label, the unmastered mix as the sum of the lanes, MIDI where the session had it. Modern music made in the DAW, pop to trance. Every creator accepted terms that name machine learning and is paid on every licence. Commercial and research licences, non-exclusive, on request. See the stems set.