Since 2 August 2026, Article 50 of Regulation (EU) 2024/1689 — what everyone calls the EU AI Act — has applied. It says two things that affect anyone generating audio.
The first is about whoever uses the system: if you generate audio that imitates a real person, or that informs on matters of public interest, you have to declare it. And if your system holds a conversation with someone, that person has to know they are talking to a machine.
The second is about whoever builds the system, which is us: synthetic content must be marked technically, in a way a machine can detect. Our adaptation deadline ends on 2 December 2026.
All audio leaving Uttera has been marked since 16 September 2026. Two and a half months ahead of the deadline, with no parameter that turns it off.
We use AudioSeal, from Meta, MIT licensed. It is an imperceptible mark embedded in the sound samples themselves, not a file metadata field: it is applied before encoding, so it travels inside the audio whatever format you ask for, and survives re-compression.
Every audio response from the API carries an X-Watermark header with the scheme used. If
you ever receive audio from us without that header, that is a bug on our side and we want
to hear about it.
The mark does not say who generated it. AudioSeal is free and public software: anyone using the same tool produces equally valid marks. So the mere presence of a mark does not prove an audio is ours, and saying otherwise would be selling you a guarantee that does not exist.
What we do prove is the opposite, which is the useful direction. The format allows 16
bits of message inside the mark, and since 19 September we use them for a fixed origin
signature: 0x5554, the letters UT in ASCII. The same for every customer and every
request.
Since there is no way to switch it off — one of our servers that could not watermark would not start — the conclusion is firm in exactly one direction:
If an audio does not carry the UT signature, we did not generate it.
And that gets used. If someone attributes to Uttera a recording you don't recognise, the absence of the signature rules it out. If someone passes off a cloned voice you never asked for and claims it came from our system, the absence of the signature disproves it.
It is a shield, not a medal. And we would rather tell it that way than sell the prettier version.
You don't have to take our word for it, or use any tool of ours. Generate any audio with
your key, install Meta's library and read the mark: if the 16-bit message equals 0x5554,
there is the signature.
⚠ One practical warning: 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 out wrong. Try a few initial offsets and keep the best reading. On WAV it does not happen.
The watermark answers "a machine made this". It does not answer "Uttera made this for you on Tuesday". That is what the PDF report is for: cryptographically signed, and verified at uttera.ai/en/verify by uploading the file — it tells you whether we issued it and whether anyone has touched it since.
Two different pieces for two different questions, and neither replaces the other.
The full detail — the Regulation's calendar, how far the mark survives, and exactly what information it carries — is in the documentation.
Anything to add or correct? Write to support@uttera.ai. If you correct us, we edit the post and credit you.
Uttera