Legal framework

Everything that governs the use of Uttera — terms of service, privacy, the data processing agreement and cookies — lives on a single page:

Legal notice and privacy →

Only the summary goes here, with a link to each section. The text that counts is the one on that page: two copies of a legal document end up diverging, and then neither is worth anything.

What it must NOT be used for

This is the section most worth reading before you integrate, because almost everything in it is a crime and not merely a breach of contract:

• Impersonating a person with a cloned voice in order to deceive someone.
• Cloning the voice of an identifiable person without their consent, even if the source material is public. Voice is protected by art. 7.6 of Spanish Organic Law 1/1982.
• Transcribing unlawfully obtained recordings. Recording a conversation you are part of is not a crime in Spain; recording one between others is (art. 197 of the Criminal Code).
• Identifying people by their voice, or building a register of voiceprints.
• Making decisions about people from tone or speaker profiling.

The full list, with the article that applies to each →

The AI Act and the audio you generate

Since 2 August 2026, article 50 of Regulation (EU) 2024/1689 applies. If you generate audio that imitates a real person, or that informs the public on matters of public interest, you must disclose it. And if your system converses with someone, that person must know they are talking to a machine.

The duty to technically mark synthetic content is ours, and the adaptation period that applies to us ends on 2 December 2026.

The timeline and the detail →

What the supervisory authority says

Spain's Data Protection Agency (AEPD) —the supervisory authority that applies to us, because the processing happens entirely in Spain— has published a piece on voice transcription with AI (in Spanish). It's worth reading before you integrate: it explains why voice is personal data, what gets inferred from it, and why retention and secondary use are the real risk. Our stance, detailed in Data privacy, aligns with that view: no retraining on your voice, no content retention, and personal-data redaction on the way out.

The watermark

All audio Uttera generates carries a watermark, since 16 September 2026. Both voices, standard and cloned, and all five formats. It is not optional and no API parameter turns it off.

We use AudioSeal, by Meta, under the MIT licence. It is an imperceptible mark embedded in the sound samples themselves, not a file metadata field. It is applied before encoding, so it travels in the audio you receive whatever the format.

Every audio response from the API carries the X-Watermark header with the scheme used. If you ever receive audio from us without that header, that is a failure on our side and we want to hear about it.

Alongside it come X-Audio-Sha256 and Content-Digest: the SHA-256 of the bytes we hand you, in hex and in the RFC 9530 form. They answer a different question from the mark. The mark says the audio was machine-generated; the digest says the file you are holding is the one that left here. Neither replaces the other, and the difference is useful: a mark can survive a re-encode, while the digest will stop matching — so if the mark is there and the digest is not, what you have has been through something since we sent it.

What it says, and what it does not

The mark says that the audio was machine-generated. And nothing else: not who requested it, nor when, nor under which account, nor any identifier that could trace it back to a person.

The AudioSeal format carries 16 bits of message inside the mark, and we use them for a fixed origin signature: 0x5554, the letters UT in ASCII. It is the same for every customer and every request.

It was needed for a concrete reason: AudioSeal is free and public software, so the mere presence of a mark says nothing about whose audio it is — anyone using the same tool produces equally valid marks.

What the signature is good for (and what it is not)

The signature does not prove that a piece of audio is ours. Because AudioSeal is public, anyone can stamp those same 16 bits. Telling you otherwise would be selling you a guarantee that does not exist.

What it does prove is the opposite, and that is the part you actually want: everything our engines produce is signed — there is no way to turn it off, and a server of ours that could not mark would not start. So if a piece of audio does not carry the UT signature, we did not generate it. That is a firm statement, and you can use it:

• if someone attributes to Uttera a recording you do not recognise, the absence of the signature rules it out;
• if someone passes off a cloned voice you never requested as yours, and claims it came from our system, the absence of the signature refutes it.

You can check this yourself, without asking us for anything, using Meta's tool: read the mark and look at the 16-bit message. If it reads 0x5554, those are the letters UT.

⚠ The signature is decoded by position, and compressed formats add a small delay at the start of the audio. If you read the mark on an mp3 without correcting that offset, the message comes back altered. Try a few small offsets at the beginning and keep the best reading.

If you need positive proof that cannot be forged, the watermark is not the right piece: use the PDF report, which we sign cryptographically and which can be checked at /verify.

How far it holds

Measured by us, on audio that came out of our engines:

We publish the case where it fails too. A mark presented as infallible is not a mark, it is advertising.

Why we do NOT use C2PA

C2PA is a signed manifest that travels inside the file. It is the other common route for marking synthetic content, and we have decided not to take it:

• One ffmpeg command strips it. It is container metadata: it goes away on re-encode, and it goes away by itself as soon as the audio passes through any platform that transcodes. It proves nothing, because anyone can remove it.
• It is not needed for compliance. Article 50(2) requires machine-readable, detectable marking. The watermark is exactly that, in all five formats. C2PA is one way to comply, not the way.
• The standard does not support opus or raw pcm, two of the five formats we serve. Verified by us against the library, not read off a documentation page.
• It does not survive the case that matters. Synthetic-voice fraud does not arrive as a file: it arrives as a phone call. There is no container there, no metadata and no manifest. There is audio, and the watermark is still in it.

What C2PA brings is interoperability with platforms that read the manifest to label content. That is real value, but it does not outweigh the above. If the facts change — if the standard starts supporting opus, if platforms begin requiring it — we will revisit it and say so here.

If you're a company sending us other people's data

You are the controller and we are the processor (art. 28 GDPR). The processing agreement is written down and serves as such; if you need a signed contract, write to support@uttera.ai and we'll sign it.

The data processing agreement →