Guides · checked 1 Nov 2026
What licensed training data means
Licensed, consented, ethically sourced, cleared: the words get used loosely. These are the questions behind them, and the answers you should get in writing from any audio data vendor, including us.
Who gave permission, and how
A creator or rights holder accepted written terms that name machine learning, or signed a consent for the set. Ask to see the clause. Ours reads: the catalogue may be used for "developing, training and improving software, audio tools and other technology, including machine learning". Every creator in every set accepted it.
Does the permission reach the company you pay
A chain: the creator grants the platform, the platform grants the licensor, the licensor signs with you. Ask who you contract with and where their right comes from. You sign with XIM Audio Inc., which holds the licence from WeMakeDanceMusic, the marketplace the creators sell on.
Recording or composition
Every song carries two rights: the recording, which is the sound, and the composition, which is the notes and the lyrics. A production-data licence covers the sound you train on: the stems, the sessions, the files. Compositions and lyrics are not licensed, and no delivery carries a song title, an artist name or a lyric. A model learns how the music was made, not a songbook.
Was any of it scraped
Audio pulled from streaming services or video sites was never licensed to the dataset by the people who own it, whatever the dataset's own licence says. Ask where the files came from. Ours are sold files from private download archives: never streamed, never crawled, never public.
Was any of it made by a model
Separated stems are a separation model's guess at the parts. Generated audio is a generative model's. Both carry the errors of the model that made them, and a dataset built from them teaches those errors as truth. Ask for a warranty. Ours: no file we license was made by a separation model or by AI.
Is anyone paid
Consent without payment does not last. Ask whether the creators share in each licence. Ours receive a share of every licence, pro rata by what of theirs is in it.
What happens when an item is pulled
Creators change their minds and claims get made. Ask what the licence says. Ours: we tell you in writing, you stop using the item in new training within 30 days and delete it, and the models you already trained are not affected.
Is it anonymised
A dataset should not carry personal data or let a buyer identify the people behind it. Ours delivers pseudonymous song and creator IDs, numbered lanes with instrument labels, and no names of any kind. Creator identity stays with the marketplace.
What do you get on paper
A licence that names the use, the term and the fee, a manifest per delivery, and a datasheet that says how the set was built: sources, dates, processing, exclusions, known limits. Our standard licence is public, and in it we promise the right to grant it, defend a third-party claim against your licensed use up to the fees you paid, and warrant that no file was separated or generated.