Guides · Article 50 · EU AI Act
AI voice synthesis tools and the EU AI Act
Voice synthesis tools — text-to-speech narrators, IVR voices, dubbing engines — form a fast-growing category, and some can also mimic a specific person. Two EU AI Act paragraphs can bear on synthetic audio: Article 50(2), the provider-side marking of generated audio, and Article 50(4), the deployer-side disclosure where cloned audio of a real person constitutes a deep fake. This guide explains both from the regulation text. It is informational only, not legal advice, and does not name any tool category as in or out of scope.
Marking generated audio — Article 50(2)
Article 50(2) asks providers of AI systems generating synthetic audio, image, video or text to ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, with solutions effective, interoperable, robust and reliable as far as technically feasible, taking into account the specificities and limitations of the content type. Audio carries its own constraints — re-encoding and compression can degrade a signal — which the paragraph expressly acknowledges. As written, this duty sits on the provider of the voice system, not on every narrator who exports a file. Whether a given category of tool falls inside 50(2) is fact-specific.
Cloned voices of real people — Article 50(4)
Where the output is audio constituting a deep fake — a convincing clone of an identifiable person's voice — Article 50(4) is written on the deployer of the system: they must disclose that the content has been artificially generated or manipulated. A generic invented narrator therefore raises a different, provider-focused marking question, while cloning a real voice raises a separate deployer-side disclosure question. The artistic-work modulation and the law-enforcement carve-out can apply. Whether your audio is a 'deep fake' and whether a carve-out applies are fact-specific.
What teams tend to separate out
- Whether the audio is synthetic at all (points toward the provider-side 50(2) marking question).
- Whether it clones an identifiable real person (points toward the deployer-side 50(4) question).
- Who is provider-side and who is deployer-side for the specific voice system.
- Timing and accessibility of any disclosure under Article 50(5).
- Counsel's review of the role and scope mapping before relying on it.
Common questions
Do voice synthesis tools need to disclose that audio is AI-generated?
The machine-readable marking duty in Article 50(2) is written on the provider of the voice system, and the deepfake disclosure in 50(4) is written on the deployer where cloned audio of a real person is involved. Which applies to your case, if either, is fact-specific — verify with qualified counsel.
See what may apply to your business
Answer seven quick questions for an automated, informational indication of which Article 50 obligations appear likely to apply — free, and not legal advice.