What the marking duty says about audio

The marking duty lists audio first and discussion of it runs almost entirely on pictures. Two entries in this record address audio, and both had to choose between a record beside the file and a signal hidden inside it. From 3 source pages, read 2026-09-12 to 2026-09-22; no maker pays to be listed.

Fig. 1 Both audio entries chose the second, which makes them the two whose marks survive a container rebuild and carry the least.

What the article asks for audio, and what the two audio entries answer. Recorded 2026-09-22.
Point	Where it sits	What the record holds
The duty covers audio	Article 50(2), provider	Synthetic audio is named alongside image, video and text
No visible layer exists	Both audio entries	A waveform has no corner to put a notice in
Coverage inside one product	One audio entry	Models from elsewhere are generally not marked

Inclusion rule. Entries whose recorded marking mechanism concerns sound. Entries describing video or images are covered on their own pages. Order. The duty first, then what the entries say about it.

1Audio removes one of the three options

For pictures a vendor can attach a manifest, embed a signal, or render a notice. For sound the third is not available in any comparable way, so the practical choice is a container record or something mixed into the audio.

Both audio entries here chose the second. That makes them the two entries whose marks survive a container rebuild, and the two whose marks carry least information.

2Speech travels through the most hostile pipelines

A generated line is mixed under music, ducked, limited, compressed for a platform and sometimes played into a room and captured again. Each step is a chance for an embedded signal to be attenuated past detection.

None of those steps appears in either entry's durability claim. One measures a quality trade-off; the other names no operations at all. The record marks the gap rather than guessing across it.

3The deployer duty is where audio gets awkward

A viewer can be shown a notice. A listener has to be told, which means a disclosure in the surrounding publication rather than in the file, and nothing in either entry helps with that.

Both entries record that nothing audible or visible is described. So for audio the publisher's half of the article is entirely the publisher's, with no tool-side assistance documented anywhere in this record.

ElevenLabs (elevenlabs.io)
Watermarking embeds imperceptible patterns into audio at frequencies masked by existing sounda speech and audio generatorElevenLabs, documentation / recorded 2026-09-22
Meta AudioSeal (ai.meta.com)
A generator and detector architecture trained jointly with a localization loss, for detection down to the sample levela speech watermarking schemeMeta AI, research publication / recorded 2026-09-22
Article 50(2)
Providers must mark synthetic audio, image, video and text in a machine-readable formatincluding general-purpose AI systemsAI Act Explorer, Article 50 / recorded 2026-09-12
ElevenLabs
Models from other providers offered through the platform are generally not watermarked unless stated otherwisecoverage inside the vendor's own productElevenLabs, documentation / recorded 2026-09-22

4A provider's own account

Against the duty in “The audio half”, SceneMixer's published pages describe what its pipeline produces:

SceneMixer writes dialogue into the script analysis, assigns each line to a shot with a speaker and a short delivery note, and the video model performs it — picture and audio generated together, so the mouth forms that sentence in the language you chose.
SceneMixer, AI Dialogue Generator (read 2026-10-13)

Dubbing replaces the audio of a finished clip; the picture was generated for a different line, so mouth shapes, pauses and shot length all belong to the original language. Native-language generation gives the video model the line in the target language before it renders the shot, so the performance, the pacing and the cut points are built for that language.
SceneMixer, Native-Language Dialogue vs. Dubbing in AI Video (2026): Why the Cast Should Speak the Line (read 2026-10-13)

5Sources

Article wording from AI Act Explorer, Article 50; vendor statements from the pages listed on each generator entry. Both read 2026-09-22. Related: Open weights, What an omission proves, Two layers at once.