Request a sample

Guides · checked 1 Nov 2026

A MUSDB18 alternative for commercial use

MUSDB18 trained a generation of separation models, and its licence stops at research. This is what a dataset has to have once the model ships, and what XIM Audio holds.

What MUSDB18 gives you

150 full songs, about ten hours, each as a mix and four stems: vocals, drums, bass and other. The mix is the sum of the stems. MUSDB18-HQ is the same set as uncompressed WAV. The set is provided for educational purposes only and may not be used for any commercial purpose without the copyright holders' permission; the audio comes from several earlier collections, each with its own terms. Every separation paper since 2018 reports on it, which is its strength and its limit.

Where it stops

Non-commercial: a product may not be trained on it. 150 songs is small for a model that has to generalise. Four buses hide the instrument detail; a hi-hat and a tom are the same stem. The genre spread is the open music of the early 2010s. And every model that scores well on it has, one way or another, seen it.

What a commercial set has to have

What XIM Audio holds

4,223 songs with every lane printed from the producer's session: 110,662 lanes, about 26 per song, about 9,153 hours, measured 4 Nov 2026. One 24-bit WAV per lane, one of 53 instrument labels per lane, and the unmastered mix of each song as the unity-gain sum of its lanes, with the residual recorded. Modern music made in the DAW, pop to trance. 2,006 vocal and instrumental pairs from the same sessions sit beside it.

Deliveries use the MoisesDB folder layout, so the moisesdb loader reads them as is, with each lane under its stem group and the producer's lane name kept. Fold the lanes into vocals, drums, bass and other and you have the MUSDB18 shape at 4,223 songs; the reverse is not possible. The package for this task ships within 10 working days of payment.

Every creator accepted terms that name machine learning and is paid on every licence. No separation model made any file. Our standard licence says so in writing, and a small evaluation set with the full index comes first. Training data for stem separation.

A few songs in the MoisesDB layout, with the full index. Ask for an evaluation set, or write to mikas@ximaudio.com. Our standard licence is public.