XIM Demix · measured 25 Oct 2026
Training data for stem separation
Separation models plateau on the few hundred public multitracks; this is thousands of real songs, every lane printed from the producer's own session, with the unity-sum mix beside it.
Each song comes with
- one WAV per lane under its stem group (MoisesDB names)
- mix.wav, the unity-gain sum of the lanes
- mix_flat_24.wav, the same sum at -0.3 dBFS
- a label, source and confidence for every lane
- data.json with BPM, key, genre, DAW, sample rate and the sum residual
Variants: 4-stem standard; electronic 6-stem (kick, bass, synths, vocals, FX, percussion); vocal 2-stem from the pairs set.
Built from
- Stems Every lane of a finished song, printed by the producer from the session it was made in.
- Vocal and instrumental pairs The vocal and the rest of the song as two aligned files from the same session.
How it ships
Readiness: ships within 10 working days of payment. Delivery to a private S3 prefix or your bucket, with a manifest, checksums, creator-level splits and a datasheet.
Loader: the MoisesDB folder layout, readable by the moisesdb loader as is.
Rights
Every creator in this set accepted terms that name AI training. Every stem is printed from the creator's own session, part by part, and the manifest says which session it came from. Read the rights page.