It's getting harder to tell whether a voice was recorded by a person or generated by a model. Watermarking is the technical answer to that question, and it's worth understanding what it actually promises, because it's easy to sell it as something it isn't.
A watermark in audio is an imperceptible signal embedded inside the sound itself that lets a machine recognise the audio was generated artificially. You don't hear it, it doesn't change what the voice says, and it isn't a written notice or a logo: it lives in the sound samples, below the threshold of what the ear perceives.
The key word is imperceptible. It doesn't spoil the audio and you can't remove it by turning up the volume or cutting out a chunk. It's woven into the signal, not stuck on top.
The temptation is to mark the file: put a tag in the mp3 headers saying "AI-generated". That's metadata, and the problem is that it falls off on its own. You re-encode the audio to another format, upload it to a social network that re-compresses it, run it through an editor — and the tag vanishes without a trace. The sound is still there; the provenance is not.
A watermark solves exactly that. Because it lives inside the samples and not in the file's wrapper, it survives re-compression, a format change, or having its start trimmed off. It travels with the sound, not with the container. That's the whole difference, and it's why a separate technique exists instead of settling for a tag.
Knowing where an audio came from has gone from a curiosity to a requirement. Since 2026 the EU AI Act obliges synthetic content to be marked in a way a machine can detect, precisely because the human route — "you can tell it's fake" — no longer works.
But beyond the law, provenance is what lets you separate two cases that get confused today: a real audio someone claims is synthetic, and a synthetic audio someone passes off as real. Without a checkable mark, that argument is your word against theirs. With one, there is something to look at.
Here's the nuance almost nobody gets right. A watermark can carry, on top of the "this is synthetic" signal, a small message: a signature. And people expect too much of the signature.
The tool we use to mark is public and free. That means anyone can produce an equally valid mark with the same software. So the mere presence of a mark — even with a signature — does not prove who generated the audio. Saying otherwise would be selling you a guarantee that doesn't exist.
What can be proven is the opposite, and that's the useful part. We embed a fixed signature in the mark, the letters UT, in all audio that leaves Uttera, with no way to switch it off: one of our servers that couldn't watermark simply wouldn't start. From that comes a conclusion that is firm in exactly one direction:
If an audio doesn't carry the UT signature, we didn't generate it.
And that gets used daily. If someone attributes to you a recording you don't recognise, the absence of the signature rules it out. If someone passes off as Uttera's a voice you never asked for, the absence of the signature disproves it.
It's a shield, not a medal. Its value isn't in saying "this is ours" — no public mark can guarantee that — but in being able to say with confidence "this is not ours". We'd rather tell it that way than sell the prettier version.
Worth being blunt, because confusion here is expensive:
A watermark is a signal hidden in the sound that survives the file's journey and lets a machine recognise generated audio. It doesn't prove authorship on its own, but a fixed signature does let you rule out with certainty what isn't yours, which is usually what really protects you. If you want to see it work — upload an audio or a report and have it tell you what's inside — it's at uttera.ai/en/verify.
Anything to add or correct? Write to support@uttera.ai. If you correct us, we edit the post and credit you.
Uttera