With AI, against AI
Services. Voices. Use cases. Compliance. Pricing.

AI voices and the EU AI Act

Approved. On record.

What the EU AI Act requires, what the release document contains and what it rules out. For you and your legal team.

  • Release document per audio file
  • Personal rights cleared
  • Ready for the EU AI Act
For your legal team

What you have in hand.

Five points your legal team will want to check. For each one, you get proof from us.

EU AI Act

Since 2 August 2026, it matters where a voice comes from.

Article 50 of the EU AI Act sets the rules for AI voices. It also applies to Swiss companies as soon as their content is used in the EU.

2 Aug 2026

Disclosure

If you use AI to recreate a real person's voice, you have to disclose it. For artistic works, a notice that doesn't disrupt the work is enough.

2 Dec 2026

Marking

Providers must mark AI audio as artificially generated in a machine-readable way. For systems already on the market before August, the grace period ends on 2 December 2026.

EUR 15m

Fines

Breaches of Article 50 cost up to EUR 15 million or 3% of worldwide annual turnover. For SMEs, the lower amount applies.

Where AI voices sit in the EU AI Act.

The EU AI Act sorts AI by risk. AI voices for advertising, e-learning or explainer videos are usually not high-risk. The transparency obligations of Article 50 apply to them.

  1. Prohibited Art. 5 For example manipulation that harms people, or exploiting vulnerabilities.
  2. High risk Art. 6, Annex III For example AI for hiring decisions, for credit scoring or to identify people biometrically, including by their voice.
  3. Usually AI voices. Transparency obligations Art. 50 AI voices and deepfakes: mark and disclose.
  4. Minimal risk For example spam filters or AI in video games. No specific obligations.

What you use an AI voice for matters. On a call, we look at your use case.

AI voices are like a fake luxury watch: the risk lies with whoever uses it.

With CrispyVoice, you get written consent and proof for every file. Crispy Watermark (beta) will add the machine-readable marking. What you need to disclose, we work out with you on a call.

As of October 2026. Regulation (EU) 2024/1689, Art. 5, 6, 50 and 99. European Commission Q&A on Article 50. Simplified. Not legal advice.

Voice and the law

For the voice itself, personality rights apply.

A voice belongs to the person

Like your own image, your voice is part of who you are. Anyone who uses it or recreates it with AI needs that person's consent. Without it, you risk injunctions and damages.

AI audio is usually not a work

Copyright protects intellectual creations by humans. Purely machine-generated audio usually doesn't qualify. So it matters little who holds "the rights to the audio". What counts is the voice artist's consent.

That's why our approval covers the voice artist's personality rights, for every file.

Simplified overview for Switzerland and the EU, as of October 2026. Not legal advice.

Case law

What courts have already decided.

Three decisions show what matters: the person's consent and clear labelling.

At CrispyVoice, every AI voice comes with the voice artist's written consent. Every file comes with a release document.

As of 29 September 2026. A selection of published decisions, simplified. Not legal advice.

Short film, 0:21

This is what it feels like when nobody asked.

Exaggerated on purpose: what happens when an AI voice goes out without the voice artist's approval. The cases above are real.

From research

“Approaches like CrispyVoice […] point in the right direction, in my view.”

Prof. Oliver Niebuhr Associate Professor of Communication and Innovation, University of Southern Denmark Interview, Meinungsbarometer.info, 25 September 2026. Translated from German. Read the interview (German).
Approval and proof

Every file with a release document.

Release document

The voice artist's name, your company, purchase date, approved purpose, reference number and a checksum (SHA-256). Any change invalidates the checksum. You show it to your legal team or your client.

Crispy WatermarkBeta

An inaudible, machine-readable marker in the audio shows that it's AI-generated and approved. Checking it takes a smartphone: point it at the voice and see the result. The video shows how it works.

How to get your licence.
✓ Valid Document
Release Document
Transfer of Utilisation & Exploitation Rights
Voice Artist
Max Sample Speaker
Buyer
Your company
Purchase Date
Automatic upon purchase
Purpose of Use
Selected medium (e.g. e-learning)
Licence
Perpetual for purchased use
Reference
cv-xxxx-xxxx-xxxx

Checksum: SHA-256 · Any change invalidates the checksum · The digital version is authoritative

The approval excludes:

  • AI training on the voice
  • Recreating the voice outside CrispyVoice
  • Passing the usage rights on to third parties
  • Hate speech, discrimination, extremist political propaganda
Licensing path

From listening to licence.

Four steps. At the end, you have the audio and the proof.

  1. Listen

    In the voice catalogue, you hear the voices in the original and as an Alter Ego.

    Go to the voice catalogue.

  2. Choose your way

    Generate it yourself in the shop, or let us produce it: with the Alter Ego or with the real voice in the studio.

    Request a quote.

  3. Set the purpose

    You decide what you'll use the voice for: in the shop when you buy, otherwise in the quote. Exactly that purpose gets approved.

  4. Audio with release document

    You get every file with a signed release document. Hand it to your legal team.

Questions

What legal teams ask.

The EU AI Act classifies AI by risk. AI voices usually fall under the transparency obligations of Article 50: anyone who realistically replicates a real person must disclose it. Whether that also applies to a voice used with consent is not yet fully settled. We recommend labelling every production. It protects you from fines of up to EUR 15 million and shows your audience you're playing fair. This is not legal advice.
No. Labelling tells your audience that an AI is speaking. Consent decides whether you may use the voice at all. You need both. With us, consent comes as a release document with every file. Whether and how you label depends on the use.
Yes, as soon as your content is used in the EU. Either way, you need the voice artist's consent, and the approval gives you exactly that.
For example: name of the voice artist, your company, date of purchase, approved purpose, reference number and a checksum (SHA-256). Any change invalidates the checksum. The digital version is authoritative.
Then you show your release document. It proves who agreed and for what purpose. Consent comes straight from the voice artist, not through a platform. With anonymous AI voices this proof is often missing, because it's unclear whose voice is in the model.
No, the usage rights are not transferable. The simplest way: your client buys. Their company is then named in the release document, and you as the agency may work with the files on that project. Want to sell voices to your clients yourself? Become a reseller. See the reseller programme.
Yes, they can, as with any right to one's own voice. That's why we settle it in writing beforehand: from the revocation on, no new files are made with their voice. What you have stays usable within your release document, advertising in a subscription until the end of the paid term plus 30 days. With a subscription, you switch to another voice right away at no cost, or you cancel without notice and get paid months refunded pro rata.
No, explicitly not. AI training and recreating the voice are excluded by contract. If you breach this, you are liable.

Let's look at your project together.

30 minutes, no obligation. Even if you don't use an AI voice yet and are just looking into it. Afterwards you'll know what to watch out for.

Und
bitte!