What SynthID is, and what it means for dubbed audio

SynthID is a watermark for AI-generated content, developed by Google DeepMind. In audio, it's embedded into the sound itself rather than attached as metadata, it's inaudible to listeners, and it's designed to survive the things that normally destroy a marker.

ElevenLabs, whose dubbing API runs behind this site, began adding it to generated audio in 2026. Their announcement, published the week of 25 June 2026 and last updated on 9 August, says they "started including SynthID in Text to Speech generations by free users" and would "expand coverage to all ElevenLabs audio generations over the coming weeks".

If you're here because you want to remove one, skip to the last section. The short answer is that you can't, and the longer answer is why you probably shouldn't want to.

How is it different from metadata?

Metadata sits alongside a file. A watermark of this kind sits inside the signal.

That distinction is the entire point. Metadata is trivially removed: re-encode the file, strip the tags, screen-record the playback, and the marker is gone while the audio is unchanged. Anyone who wanted to hide a file's origin could do it in one command.

SynthID is embedded in the audio itself, as modifications too small for a listener to perceive but detectable by a model that knows what to look for. Removing it would mean damaging the audio.

ElevenLabs states the watermarks "remain even when clips are trimmed, sped up, stripped of metadata, or converted into a different file type", and says they hold up against cropping and other transformations commonly encountered online.

So the usual tricks don't apply. Converting WAV to MP3 doesn't clear it. Nor does trimming the first ten seconds, changing the speed, or stripping every tag in the file.

Who can read it?

Anyone, which is the part people don't expect.

ElevenLabs publishes a free Audio Detector at elevenlabs.io/app/audio-detector. It checks for the watermark first, then falls back to their older AI Speech Classifier when it doesn't find one. You don't need an account relationship with whoever generated the file, and you don't need the original.

That's a genuine shift in the provenance question. Detection used to depend on a listener noticing something odd, or on a classifier guessing from acoustic features, and both are unreliable and getting worse as models improve. A watermark is a different kind of evidence: either it's there or it isn't.

What it doesn't do

Four limits worth stating, because the coverage tends to oversell it.

It doesn't identify you. The watermark marks audio as generated by a particular system. It isn't a signature tying a file to an account, and reading it doesn't tell the reader who made it.

It doesn't cover audio from other systems. It's per-provider. A detector built for one vendor's watermark says nothing useful about a file generated somewhere else, and an unwatermarked file is not evidence of a human recording. It's just an absence.

It doesn't survive everything. Surviving ordinary handling is not the same as being indestructible. Any watermark can be degraded by sufficiently aggressive processing, and the practical question is whether the processing that removes it also ruins the audio. That's the design target rather than a guarantee.

It doesn't tell you whether a dub is allowed. Provenance is not permission. A watermarked dub of your own video is fine. An unwatermarked dub of someone else's copyrighted material is still infringement. The marker describes how something was made, not whether you were entitled to make it.

Why is this happening now?

Because it's being legislated into existence, and the timing isn't a coincidence.

Article 50 of the EU AI Act became applicable on 2 August 2026. It requires providers of AI systems that generate synthetic audio, image, video or text to ensure that output is "marked in a machine-readable format and detectable as artificially generated or manipulated".

Read that as a specification and you get roughly what SynthID is: a marker a machine can find, attached to the content rather than to a file wrapper, resilient enough to survive ordinary handling.

That obligation falls on providers of the AI system, not on you as a creator using one. Your obligations are different and mostly narrower, and we've covered when you actually have to label an AI-dubbed video, including the separate question of what YouTube requires.

Does dubbed audio from this service carry it?

We're not going to answer that with a claim, because we can't source one.

ElevenLabs' published rollout post doesn't name Dubbing among the products covered at the time of writing, and their help centre article describing current coverage isn't publicly readable. Any statement we made in either direction would be a guess presented as a fact, and this is exactly the kind of question where that does real damage to someone relying on it.

What we can tell you is how to find out for your own file: run it through ElevenLabs' Audio Detector. It's free, it takes a minute, and it answers the question about the actual file you have.

Our wider post on whether anyone can tell your video was dubbed by AI separates the three questions people ask as one, since a listener, a platform and a detector give three different answers.

For what it's worth on the input side: we never take a voice sample. There's no enrolment step and no stored voice profile, because the upload endpoint takes a file and one of 32 target languages and nothing else. You don't declare the source language either, since it's detected from the audio.

Can you remove an AI audio watermark?

No, and we won't help you try.

The processing that would degrade a watermark of this kind also degrades the audio, which defeats the purpose. And the useful question underneath is why you'd want to.

If you're dubbing your own content, a watermark costs you nothing. It marks the audio as AI-generated, which is true and which nobody is hiding. YouTube explicitly exempts dubbing your own voice from disclosure, and the EU deep fake obligation turns on content that "would falsely appear to a person to be authentic", which a translation of your own words does not.

If removing it would help you, the thing it would be helping with is passing off synthetic audio as something it isn't. That's the case the watermark exists for, and it's not one we're going to write a workaround for.

Where to read next

If you're planning dubbing work, the per-language guides cover the practical side: dubbing YouTube videos from English to Korean is one of the pairs where synthetic voice quality is scrutinised most, dubbing marketing video from English to Portuguese is another, and the use case index has the rest.

FAQ

What is SynthID?

A watermarking technology from Google DeepMind that embeds an imperceptible marker into AI-generated content. In audio it lives in the signal rather than in metadata, so it survives conversion and editing that would remove a tag.

Does ElevenLabs watermark audio?

ElevenLabs began including SynthID in generated audio in 2026, starting with Text to Speech for free users the week of 25 June and stating it would expand to all their audio generations. Their published post does not name Dubbing specifically, so check a given file with their free Audio Detector rather than assuming either way.

Can you remove an AI audio watermark?

Not through normal editing. ElevenLabs states the watermark survives trimming, speed changes, metadata stripping and format conversion. Processing aggressive enough to threaten it would also degrade the audio.

Does a watermark mean my dub is not allowed?

No. A watermark records how audio was made, not whether you had the right to make it. Dubbing your own content is fine and stays fine whether or not the output is marked.