AI Stem Separation and Copyright: Can You Release Stems From Someone Else's Song?

A producer sends you an isolated vocal from a hit record. The hook fits the track you are building, and the file sounds clean enough to drop straight into the session. Before you arrange around it, ask where the audio came from and who cleared its use. An AI separator answers neither question.

The file changed; the rights did not

Stem separation can pull a vocal, drum part, bass line, or other layer out of a finished mix. The software changes what you hear in the session. It doesn't transfer ownership, grant a sample license, or tell you that the source file was cleared for a commercial release.

A separated part can still contain the original performance and recording. The vocal may be easier to hear on its own, but it came from the same sound recording. If you build your chorus around that take, the source has followed the stem into your new session.

A clean separation can feel like a new asset because it arrives as a new file. Treat it as a production shortcut, not a clearance shortcut. If another producer hands you the stem, ask for the original song title, the source file, the exact separation tool, and the permission that covers the new use. Keep that record with the session instead of trying to reconstruct the story after the mix is finished.

A streaming subscription doesn't settle the question either. Listening access is different from permission to extract or alter a track. Moises' help center says its stem-separation service requires audio you own or have legal permission to use, and that streaming access doesn't give you permission to modify the recording (Moises help center).

One song can involve two different copyrights

The U.S. Copyright Office describes a musical work and a sound recording as separate copyrighted works. The musical work is the composition and lyrics; the sound recording is a particular fixed performance. They can have different owners and licensing routes (U.S. Copyright Office, What Musicians Should Know about Copyright).

An AI-isolated vocal usually keeps actual audio from the recording. That makes it closer to sampling than interpolation. The Copyright Office defines sampling as taking part of an existing sound recording and incorporating it into a new work. Interpolation uses part of the underlying musical work but records new audio instead. For a sample, licenses from both the sound-recording owner and the musical-work owner may be necessary; an interpolation can still require permission for the composition (Copyright Office sampling and interpolation guide).

If a producer reconstructs a melody with a new performance, that changes which layer is being used. It doesn't make the composition disappear. If the original isolated audio remains in the bounce, the sound recording is still part of the conversation.

This is why a single yes from one contact may be incomplete. The label or artist who controls a master might not control the song's composition, and a publisher's permission to use a composition doesn't automatically license a particular studio recording. Ask who controls each part before you agree to a fee, a credit, or a release date.

A fair clearance map starts with two lines: what you can do with the sound recording, and what you can do with the composition. Record the song title, source version, rights contact, permitted use, release platforms, edits, territory, and any fee. If you can't identify one owner, don't treat a green light from the other as a complete yes.

Check the tool before you upload

A license to use a software feature and permission to use the source song are separate issues. Read the specific terms for the exact feature you plan to use, not just the marketing page or subscription tier.

For example, Moises' Terms put its Real-Time Stem Separation feature under a feature-specific rule: it is for personal use, users must have the rights or permissions needed to process the audio, and the terms prohibit recording, storing, exporting, or redistributing the separated audio from that feature (Moises Terms, Section 12.3). Those are Moises' rules for that feature. They shouldn't be treated as universal terms for every separation tool, and they don't grant rights in the song itself.

There are two questions to answer before you start: are you allowed to give this audio to the tool, and are you allowed to use the resulting material in the way you intend? A yes to one doesn't automatically answer the other. A practice session, a remix contest, a paid release, and a video sync can all sit under different permissions.

Clear the use before the final mix

Before a third-party stem becomes part of a release, pin down the source and intended use. A short email saying "sure, use it" can leave the hard questions open: which recording, which version of the composition, what edits, what release, and what platforms or territory?

The best time to trace this is during arrangement, while replacing the part is still easy. If permission only covers a private practice, contest entry, or a short promotional clip, don't assume it covers a commercial release. If no answer arrives, build a new original part instead of forcing the deadline. Re-recording can avoid using that particular sound recording, but a recognizable melody or lyrics can still involve the composition.

• Start with your source file. Keep the license, email, or official stem-pack terms that show what you may do with it.
• Identify both sides of the song: the sound-recording owner and the musical-work owner or publisher. They may be different parties.
• Describe the use you want cleared, including the release, edits, commercial use, distribution, and any video or live-performance plans.
• If you replay the phrase instead of using the isolated audio, clear the composition rights that still apply to the new performance.

A label, publisher, distributor, or artist may point you to the right contact, but a stream count, credit line, or software subscription doesn't replace the permission itself. Save the actual license or written approval with the session files so the next collaborator can see what the stem is cleared to do.

Fair use can apply in some U.S. situations, but it is fact-specific. The Copyright Office says there is no preset number of seconds or percentage that makes a sample automatically safe; courts assess the circumstances of each use (U.S. Copyright Office sampling guide). Don't build a release plan around a supposed "short enough" rule.

The cleanest stem in your session still needs a clean path to release.