ElevenLabs on a pattern hidden beneath existing sound
The mechanism described is an embedded signal rather than an attached record, placed where existing sound hides it. The marking paragraph covers synthetic audio in the same breath as image and video, and this is the audio answer to it. As of 2026-09-22.
| Point | What the record holds |
|---|---|
| Article 50(2), machine-readable marking | Imperceptible patterns embedded at frequencies masked by existing sound |
| Read from | Product documentation, written for somebody integrating it |
| What it does not settle | What survives a re-encode, a room, or a second capture |
Inclusion rule. One cell of this generator's entry, quoted from the page the entry names. Coverage, commentary and third-hand summaries are not admitted, in either direction. Order. Fixed order: the cell, the page it came off, then the limit on reading it.
1The paragraph names audio first, and most of the record forgets it
The marking duty lists synthetic audio, image, video and text. Discussion of it runs almost entirely on pictures, partly because pictures are what people share and partly because a watermark in a corner is easy to imagine. Audio has no corner.
That makes the mechanism here interesting rather than exotic. There is nowhere to put a visible notice in a waveform, so the only options are a record attached to the file and a signal hidden inside the sound. This entry describes the second.
2Masking is the trick that makes it inaudible
Hiding a signal under sound that is already there is the same idea lossy audio compression uses to throw material away: the ear does not notice what louder neighbouring sound covers. Placing a mark in those gaps buys imperceptibility without spending loudness on it.
It also ties the mark's fate to the content around it. Quiet passages offer less cover than dense ones, and a processing chain that rewrites the spectrum is rewriting the place the mark lives. The page does not go into that, and the cell does not invent a figure for it.
3Speech travels through more hostile pipelines than pictures do
A generated voice line rarely reaches an audience as delivered. It is mixed under music, compressed for a platform, sometimes played into a room and recorded again. Each of those is a chance for an embedded signal to be attenuated past detection.
None of that is a criticism of the mechanism, which is the only one available for this medium. It is a reason the checking question matters more here than elsewhere, and the reach of the detector is recorded on its own row for that reason.
- ElevenLabs (elevenlabs.io)Watermarking embeds imperceptible patterns into audio at frequencies masked by existing sounda speech and audio generator
- Article 50(2)Providers must mark synthetic audio, image, video and text in a machine-readable formatincluding general-purpose AI systems
4Sources
Wording taken from the documentation at elevenlabs.io, consulted 2026-09-22. All four rows for this generator sit on ElevenLabs; the same column across every entry is read down on What it asks for. Filed beside it: AudioSeal on marking speech, Nova Reel on two marking layers.