SenseDefend

Synthetic-media instruments, by article and effective date

AudioSeal on buying detection without paying in audible quality

Every audio watermark spends something. A stronger mark is easier to detect and easier to hear, and this entry's claim is about how that exchange was made rather than about surviving any particular edit. As of 2026-09-22.

What a generated voice line goes through after deliveryA picture is cropped and recompressed. A generated voice line is mixed under music, ducked, limited, compressed for a platform, and sometimes played into a room and captured again. None of those appear in the publication's list.The markPlaced where louder sound masksitStrength depends on theaudio around itInaudible, per thepublicationThe mixMusic, effects, ducking,limitingRewrites the spectrum themark lives inNot addressed on the pageThe platformLossy encoding for deliveryDiscards quiet detail bydesignNot addressed on the pageAudio an audience hearsWhere the recorded claim stops
Fig. 1 What the entry records is a quality trade-off, which is the axis the publication actually measured.
Meta AudioSeal on the what survives column, and what that column leaves open. Recorded 2026-09-22.
PointWhat the record holds
What survivesImperceptibility, claimed from a masking-based training loss
Read fromA research publication describing the method
What it does not settleWhat survives a mix, a platform pass, or a room

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.

1Masking is the reason a mark can be inaudible at all

The ear does not notice quiet detail next to louder sound, which is why lossy audio compression can discard so much. A watermark placed in those same gaps buys cover from the content rather than from loudness.

Building that principle into the training loss rather than applying it afterwards is the method's claim. It ties the mark's strength to the material, which is a property worth knowing when the material is quiet dialogue.

2Durability for audio means something different than for pictures

A picture is cropped and recompressed. A generated voice line is mixed under music, ducked, limited, compressed for a platform, and sometimes played into a room and captured again.

None of those are in the paper's list, and the cell does not add them. What it records is a quality trade-off, which is the axis the publication actually measured.

3Why this sits under persistence rather than under marking

The scheme's architecture is recorded on the marking row of the same entry. This row is about whether the mark stays worth having, which for audio is mostly a question of whether it stayed inaudible while staying findable.

Keeping them apart avoids a common compression: a paper that describes both an architecture and a trade-off is often summarised as making one claim, and the two have different strengths of evidence behind them.

  • 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(7)
    The Commission is to encourage codes of practice on detection, marking and labelling, and may adopt an implementing act if a code is inadequateprocedure in Article 98(2)AI 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 survives. Filed beside it: Synthesia on a mark that stays, Hailuo on an unmodified request.