Compliance
Speaker consent and licensing in AI speech datasets
What a defensible consent form actually says, how consent maps to files, and what licence terms let you train, distribute and commercialise without re-negotiation.
- Consent
- Written, per participant
- Licence
- Perpetual, transferable
- IP
- Transfers on delivery
- Audit
- Consent index supplied
The short version
- Consent names AI model training, derivative models, redistribution of trained models and commercial use — vague research-only wording blocks release later
- Compensation is disclosed to the participant and recorded
- Consent is signed before recording, never batch-collected afterwards
- A consent index maps participant to speaker ID to file IDs, so any file can be traced to its authorisation
- Voice-cloning and synthetic-voice use is a separate, explicit consent clause when a TTS build requires it
Why it matters in practice
Most legal blockers in AI data procurement appear at model-release time, not at purchase. They are almost always caused by consent language that covered research but not commercial deployment, or by an inability to prove which recordings came from whom.
Both are cheap to prevent at intake and very expensive to fix after a corpus is delivered.
How it is operated
| Clause | Why it matters |
|---|---|
| Purpose: AI training | Without it, the corpus is unusable for its actual purpose |
| Derivative models | Models trained on the data are derivatives; silence here creates ambiguity |
| Commercial deployment | The single most common gap that blocks a launch |
| Redistribution | Needed if the buyer licenses the dataset onward or open-sources a model |
| Synthetic voice | Separate consent required for voice cloning or TTS persona builds |
| Withdrawal | Defines what happens operationally, not just in principle |
What you receive
- The relevant policy or template as a document, not a claim on a web page
- Programme-specific artefacts delivered with the corpus manifest
- Named contact for audit questions during and after the programme
- Written confirmation at close-out
If your legal or procurement team has a questionnaire, send it with the specification. It is faster to answer it once, up front, than to unblock a signed programme later.
Frequently asked
Do speakers get paid?
Yes. Participants are compensated at fair local rates, the amount is disclosed on the consent form, and payment is recorded against the participant.
Can we open-source a model trained on this data?
Yes, when redistribution is included in the consent. It is included by default in our template; say so up front and it stays in.
What about voice cloning?
Voice cloning and synthetic persona use require a separate, explicit clause and typically a higher rate. It is never assumed under a general training consent.
Can we audit consent?
Yes. The consent index and sample signed forms are available for audit under NDA at any point during or after the programme.
Related pages
Send your compliance questionnaire with the spec
We answer procurement, legal and security questionnaires alongside the technical scope, in the same working day where we can.