SenseDefend

Synthetic-media instruments, by article and effective date

AudioSeal and SynthID on how precisely a mark can be located

Both describe embedded signals and they claim different properties. One is about resolution, the other about endurance, and a reader who wants both has to hold two pages at once. As of 2026-09-12.

Resolution and endurance are separate axesA mark can be findable in a whole clip and useless for saying which seconds are generated. It can also be precise on a clean file and gone after a platform pass. Neither property implies the other, and each entry claims one.Claims resolutionClaims enduranceQuestion answeredWhich part of this was generatedIs the mark still hereMeasured onThe material in the publicationFour named routine operationsWhat it does not claimSurvival through a mixAny position inside the fileKind of pageA research publicationA research and product pageAn assembled timeline sits in the gap between them
Fig. 1 Between them the pair shows what a complete claim would look like, and that nothing in this record makes one.
Meta AudioSeal and Google SynthID on the same column, point by point. Recorded 2026-09-22 and 2026-09-12.
PointMeta AudioSealGoogle SynthID
Property claimedLocalized detection, down to the sample levelSurvival through cropping, filters, rate changes, compression
Question it answersWhich part of this file was generatedIs the mark still here after the route
Kind of page it was read fromA research publicationA research and product page

Inclusion rule. Two entries whose published answers to one column pull in different directions. Cells hold each vendor's own wording from the page its entry names; pairs that agree on the column are not filed at all. Order. Fixed order, and the two columns are in the order the heading gives them.

1Resolution and endurance are separate axes

A mark can be findable in a whole clip and useless for saying which seconds are generated. It can also be precise on a clean file and gone after a platform pass. Neither property implies the other.

Because these two entries each claim one, the pair shows what a complete claim would look like and that nothing in this record makes one. Each page measured the axis its authors cared about.

2Assembled timelines are the case nobody claims

Almost every delivered asset is built from marked and unmarked material. A whole-file verdict cannot describe that, and a localization claim measured on clean audio does not promise to hold after a mix.

So the interesting case sits in the gap between these two pages. The record notes the gap rather than filling it, because filling it would mean inventing a measurement nobody published.

3Two research-flavoured entries, two different statuses

One is a method anybody may implement, with no claim about which products use it. The other is described as deployed across a family of consumer products, with no product names.

Neither tells a reader whether a specific file they hold was marked. That is a property of a pipeline, and the closest a register can come is recording what each page said and when it was read.

  • 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
  • SynthID
    Designed to stand up to cropping, filters, frame-rate changes and lossy compressionstated robustnessGoogle DeepMind, SynthID / recorded 2026-09-12

4Sources

Read from the research publication at ai.meta.com and the SynthID page at deepmind.google, consulted 2026-09-22 and 2026-09-12. Full entries: Meta AudioSeal and Google SynthID. The column itself is on What it asks for. Other pairs on file: Wan and Firefly, SceneMixer and Pika.