Guides · checked 1 Nov 2026
Before you license audio data
The questions audio AI teams ask before they license training data, answered in plain words with the sources named. Written from the data side, kept current as the catalogue moves.
- Open music datasets and what their licences allowMUSDB18, MoisesDB, MedleyDB, Slakh, FMA and more: each one's licence and whether a commercial model may train on it.
- A MUSDB18 alternative for commercial useWhat a separation dataset has to have once the model ships, and where to find one.
- What licensed training data meansConsent, chain of rights, recording versus composition, scraping: the questions to ask any audio data vendor.
- Stems from the session, or separated stemsTwo files can both be called a stem. Why the difference decides what a model learns, and how to tell them apart.
- DAW project file datasets: what exists, what a session carriesNo open corpus of Ableton, Logic or FL Studio sessions exists. What can be downloaded, what a session file holds, and where licensed sessions come from.
- MIDI datasets and what their licences allowLakh, MAESTRO, Groove, GigaMIDI, Aria, PDMX and more: where each file came from, its licence, and whether a commercial model may train on it.
Looking for a set? Training data by task or every dataset.