Audio data for AI training
The inside of the song, licensed for AI training.
Real DAW sessions, the stems printed from them, the MIDI and the synth presets, from working producers who authorised machine learning in writing. Measured, documented and licensed one set at a time.
9 datasets · 57,004 stems · 12,426 sessions · measured 17 Oct 2026
The lanes of one WeMakeDanceMusic session, : tracks as the producer left them, each printed from the project on its own, never separated. Press play, then mute or solo any lane, as many as you like.
Why this data
Ground truth, not model output.
A pre-authorised, documented, non-public corpus of how music is made: the sessions, stems, MIDI and presets of real tracks, pop to house to hip hop, from producers who sell on WeMakeDanceMusic.
Authorised in writing
Every creator has accepted terms that name machine learning. Nothing here needs a consent process.
Not on the internet
Sold files from private download archives. Never streamed, never crawled.
Whole sessions, not snippets
Each session is a producer's full decision chain, from the first kick to the master.
No separated stems, ever
Every stem is a lane of the session, printed from the project. No separation model made any file we license.
What we license
Nine datasets from one catalogue.
Measured and described one by one, with the rights stated per set. Every number on this site comes from the live catalogue.
Listen
Hear the data.
The player at the top of this page plays the lanes of one studio session, kick, clap, each hat, each pad, each vocal take, in the order the producer left them. Each lane was printed from the project on its own. Take one out and the others stay exact, because they were never mixed together.
What that proves
A separation model guesses at the parts of a finished mix. These files are the parts. The drums you hear are the drum lane of the session, with the producer's plugins loaded, nothing bled in from the bass or the synths.
Every stem we license was made the same way, from the producer's own session, part by part, and the manifest names the session each one came from.
What it trains
Name the task. The sets are picked for you.
- Stem separationwith parts that were never mixed, paired with their unmastered mix.
- Generative modelsthat learn structure, not only sound.
- DAW assistantsthat arrange, mix and route like a producer.
- Mixing and masteringfrom channel settings and the finished master.
- Transcription and sound designfrom MIDI and presets beside their own audio.
How a delivery works
One request, one evaluation set, one bucket.
- Pick the sets
Add the datasets you need to a request and tell us what you are building. One form, one email back from us, no automated replies.
- Run the evaluation set
A few songs with the session, the stems, the MIDI, the presets and the master, with the labels and the per-asset records. Access the same week.
- Take delivery
A private S3 prefix or your own bucket, one scoped key per deal, a manifest and a datasheet with every set. WAV as sold, FLAC on request.
Where it comes from
Real sessions, not scraped audio.
Every file comes from the WeMakeDanceMusic catalogue: Logic, Ableton, FL Studio and Cubase sessions that producers built and sold there. The stems are exact because they were never mixed in the first place. The session, the stems, the MIDI and the presets of one track sit side by side.
Creators are paid on every licence. Every creator accepted terms that name machine learning training. Rights are stated per dataset on the rights page. Nearly every product carries its BPM, key and genre, set by the producer who made it.
About
Made by producers.
XIM Audio licenses audio data for AI training. It is founded by Mikas, founder of wemakedancemusic.com, a marketplace where producers have bought and sold project files, stems and samples since 2012.
XIM is short for mix, backwards. The parts before the mix are what we keep.
Questions, labels, partners: mikas@ximaudio.com