Skip to content

Stems & AI Training

When you upload stems, FreeMix offers an optional checkbox: “Help improve FreeMix’s stem separation.” Checking it gives your consent to use your stems for separation model training. This page explains exactly what that means.

This consent is off by default, completely optional, and never affects anything else about your account, your uploads, or what you can sell.

FreeMix’s stem separation feature uses the best open source stem separation models that run on our own infrastructure. These provide excellent separation for the four stem types we offer now: drums, bass, vocals, and “other” which is everything else. Yet, we can see a world where music creators will want more stem types: synth, keys, guitar, horns, percussion, and more.

We’re building our own models for these specialized stem types to make stem separation better for everyone. These models improve when they learn from real, high-quality, isolated stems — exactly the type of stems creators upload to FreeMix.

By giving your consent, you’re allowing us to use your uploaded stems to improve the quality of these models — nothing more. It is not a grant to train general-purpose AI, generate music, or anything beyond making FreeMix’s own stem separation better. You’ll be helping the community grow by improving the tools that everyone on FreeMix can use. That’s what FreeMix is all about.

There’s another reason this matters: most stem-separation tools improve by sending your audio off to a third-party service. We’re building our own models specifically so your audio never has to leave the platform. Opting in is how that becomes possible — a better tool, built without shipping anyone’s music to outside companies.

Opting in is narrow on purpose. FreeMix will never:

  • Sell your audio to anyone, for any purpose.
  • Use your stems for anything other than training its own separation models — no generative AI, no licensing to third parties, no resale.
  • Share your raw stems with outside companies or services.

Model training happens on FreeMix’s own infrastructure. Your audio stays protected.

Training data isn’t tied to your identity. No personal information or metadata is stored alongside the audio or shared as part of the training process. The models learn by analyzing sound patterns, not who made it or where it’s from.

We want to be as transparent as possible here.

Consent is permanent. It works the same way as a FreeMix license: once you grant it, you can’t revoke it. This is because of how model training actually works.

What you can do: if you change your mind, you can email us to request that your stems be excluded from future training runs. We’ll honor that. It means your audio won’t be used to train new versions of the models going forward — but it can’t undo training that has already happened.

Put simply:

  • Grant: permanent, like a license.
  • Future training: you can opt out of future training by request, anytime.
  • Past training: cannot be reversed, even if your stems are deleted.

To request exclusion from future training, email privacy@freemix.fm.

The checkbox appears on the upload form only when your upload includes stems — either stems you’re uploading directly, or stems FreeMix generates for you through automatic separation. If your upload has no stems, you won’t see it.

  • Leave it unchecked (the default) — your stems are never used for training.
  • Check it — you consent to the terms on this page for the stems in that upload.

Your choice applies per upload. There’s no reward for opting in and no penalty for declining — it’s genuinely up to you.

If anything here is unclear, or you want to request exclusion from future training, reach out at privacy@freemix.fm. For the formal terms, see the Terms of Service.