SenseDefend

Synthetic-media instruments, by article and effective date

AudioSeal on training the generator and the detector together

This entry is a published scheme rather than a product feature, and it is the only one here whose marking row comes from a research paper. Its claim is about granularity: not whether a clip was generated, but which part of it was. As of 2026-09-22.

Whether it was generated, or which part of it wasA detector can report a verdict for a whole clip, or report where in the clip the mark is present. The second is what a localization term in the training loss buys, and it is the question a mixed timeline actually raises.Whole-clip verdictLocalized detectionWhat it reportsThis file contains a markThese seconds carry the markOn a mixed timelineOne answer for everythingAn answer per regionUseful forScreening a libraryDescribing a specific deliverableWhat it still cannot sayWhich model, at what settingsWhich model, at what settingsGranularity is this entry's claim
Fig. 1 Almost every delivered asset is assembled from marked and unmarked material, which is the case a whole-clip verdict cannot describe.
Meta AudioSeal on the machine-readable marking column, and what that column leaves open. Recorded 2026-09-22.
PointWhat the record holds
Article 50(2), machine-readable markingA generator and detector trained jointly, with a localization loss
Read fromA research publication, written to be reproduced and checked
What it does not settleWhich shipping products, if any, apply the scheme to their output

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.

1Training the two halves together changes what the detector can do

A watermark and its detector can be designed separately, with the detector taught afterwards to recognise whatever the embedder produces. Training them jointly lets the training target include properties the pair should have together, which is how a localization term ends up in the loss at all.

The consequence is a detector that reports where rather than only whether. For a transparency question that is the difference between knowing a file has generated audio in it and knowing which seconds of a mixed timeline are generated.

2Detectable as generated does not say how much of it

The paragraph asks for output marked so that it is detectable as artificially generated or manipulated. Read strictly that is a property of an output, and an output is whatever the system handed back. A clip assembled afterwards from marked and unmarked material is not an output of anything.

Sample-level localisation is the closest thing in this record to an answer for that assembled case, which is the case almost every production actually delivers. The record notes the capability and does not claim the clause requires it.

3A scheme and a product are different kinds of entry

Every other machine-readable cell here was read off a page describing something a customer can buy or call. This one was read off a paper describing something anybody can implement, which is a weaker statement about the world and a stronger one about what is possible.

The register keeps it because the marking duty is addressed to providers, and a published scheme is what a provider reaches for when it decides to meet the duty. What no page here establishes is which providers did.

  • 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)
    Technical solutions must be effective, interoperable, robust and reliable as far as technically feasiblequalifier on the marking dutyAI Act Explorer, Article 50 / recorded 2026-09-12

4Sources

Wording taken from the research publication at ai.meta.com, consulted 2026-09-22. All four rows for this generator sit on Meta AudioSeal; the same column across every entry is read down on What it asks for. Filed beside it: Nova Reel on two marking layers, Designer on credentials for images.