Speaker releases: what consent do voice data buyers expect?

Filed under Consent. Written by Sell Voice Data, November 7, 2026. About 4 minutes.

A voice recording carries two sets of rights. One belongs to whoever made and owns the recording. The other belongs to the person speaking, whose voice is part of who they are. A buyer licensing audio to train a model needs both cleared, and the second set is where most voice collections fall short.

Why do buyers care so much about consent?

A model trained on someone’s voice can learn to recognize them or reproduce them. That is a different use from playing an episode to listeners, and speakers who agreed to the first have not necessarily agreed to the second. Buyers carry the risk if a speaker later objects, so they push it back to the seller: show us the paper, or we cannot take the file.

Volume does not help here. A thousand hours with one unreleased guest in every episode can be harder to sell than a hundred clean hours from a single narrator under contract.

What should the release actually say?

There is no single standard form, and your own lawyer should write or check yours. Buyers tend to look for the same handful of points.

The use is named

The release says the recordings may be licensed to third parties to train, test and improve AI systems. “Any media now known or later developed” was common in older releases; some lawyers read it broadly, others do not, and a buyer will not want to argue the point.

Every speaker is covered

Hosts, co-hosts, guests, narrators, callers and anyone audible in the background. A podcast with a different guest every week can mean hundreds of releases.

Identity and cloning are addressed

The release should say whether the speaker may be identified, whether their voice may be used to generate new speech, and whether either is ruled out. Some speakers will agree to recognition training and refuse synthesis; you need to know which recordings carry which permission.

Payment and withdrawal are clear

What the speaker receives, if anything, and whether they can withdraw later. A buyer will ask what happens to its license if a speaker withdraws.

How do releases differ by type of recording?

The audiobook row catches people out. A narrator reading a novel is performing someone else’s words, so even with the narrator’s release, the text may still need clearing with the author or publisher.

Is a voice biometric data under US law?

It can be. A voiceprint is listed as a biometric identifier in Illinois’s Biometric Information Privacy Act, which requires informed written consent before one is collected and lets individuals sue. Texas and Washington have their own biometric identifier laws. Whether a given dataset involves voiceprints depends on what the buyer does with it, so ask your lawyer and ask the buyer.

Separately, every state sets a rule on recording conversations. Some need one party’s consent and about a dozen need everyone’s. Check your state’s statute. That rule decides whether the recording was lawful to make in the first place.

What about “this call may be recorded”?

That notice is about making a recording for a stated purpose such as quality or training of staff. It is not, by itself, consent to sell the recording to an AI company later. If your recordings include members of the public who heard only that notice, expect buyers to treat them with caution. Paid voice talent who read IVR prompts under contract are a far cleaner case.

Can you fix missing releases?

Usually, yes, at least in part.

1. List every speaker in the archive and the hours each contributes.
2. Sort them into signed, reachable and unreachable.
3. Send reachable speakers a short, plain release that names AI training and says what they will receive.
4. Drop the recordings of anyone who declines or cannot be found.

The set gets smaller, but what remains can be sold with confidence. Many sellers find that a handful of regular voices account for most of the hours anyway.

What else will a buyer check?

Ownership comes first: the recordings have to be yours, or you need written permission to license them. Client sessions are normally the client’s property unless your contract says otherwise. Personal details come next. Buyers generally de-identify what they receive, but you should still strip names, contact details and anything a speaker said about third parties before sharing even a sample. And remember that any estimate you see, on this site or elsewhere, is not an offer.

If you are weighing whether to sell voice data at all, get the releases sorted first: every later conversation goes faster with them in hand. The speaker releases checklist is a one-page version of this post.

Paper first, audio second: that order saves months.