Skip to content

Deepfake Total

Professional

Vendor: Status uncertain as of 2026-05 | Type: Detector | Cost: Contact vendor

TL;DR

A stub card. Deepfake Total emerged from the DW Innovation November 2025 audit (DW Innovation audio detection audit) as the strongest of the three audited audio detectors at 7-of-10, but the toolkit's research base did not produce an independent deployment evaluation sufficient for a full tool card. Use as the audit-leader reference; route through DW Innovation audio detection audit for the source audit and through Hiya for the press-button SEA option via InVID-WeVerify.

Why this is a stub

This card is deliberately short. The toolkit's editorial position is that an audio-deepfake-detector card requires (a) verbatim limitations from independent audit evidence, (b) a Country and platform applicability section grounded in documented SEA deployment, and © a privacy and threat model section that addresses cross-border data flow specifically. The toolkit's evidence base supplies (a) through DW Innovation audio detection audit (DW Innovation, November 2025) but does not supply (b) or © at the level the toolkit's reference cards demand. Instead of improvising the missing material, the toolkit ships this stub and routes the reader explicitly to the neighbouring tools that ARE sufficiently documented.

This stub framing is a feature: the toolkit acknowledges limited coverage transparently instead of padding a card with inferred content.

Independent accuracy

Vendor claim vs independent assessment

Vendor claim: see vendor docs; no sharp vendor headline accuracy figure is on record.

Independent finding: DW Innovation, Synthetic Audio Detectors Put to the Test (September 2025; the November 2025 audit referenced as DW Innovation audio detection audit ). Deepfake Total correctly identified 7 of 10 cases on the audit's 10-sample multilingual dataset: the strongest of the three audited audio detectors (Hiya: 4 of 10; DeepFake-O-Meter: model-to-model contradictions severe enough to undermine operational use). The audit's overall finding is that none of the three audited audio detectors "reliably identify AI voices across languages."

Limitations

Limitations

  • DW Innovation Nov 2025 audit conclusion: tools tested do not "reliably identify AI voices across languages".
  • 10-sample dataset gives qualitative ranking, not a statistically powered benchmark; no per-language or Asian-language breakdown was published.
  • Coverage and access conditions on the vendor side not independently documented; this is the reason the card is a stub rather than a full entry. Verify directly before any operational use.

How to access: route through

For audio-deepfake-detection work in the toolkit's workflow, route through the documented neighbours rather than relying on this stub:

  • Press-button SEA option: Hiya / Loccus via the InVID-WeVerify voice-clone tab. Hiya is the only press-button SEA-accessible audio detector in the shortlist; carries the DW Innovation 4-of-10 finding verbatim.
  • Institutional escalation: TrueMedia.org (Georgetown McCourt School), in closed beta as of May 2026; ensemble approach with mission-aligned radical-transparency framing.
  • Audit-source reference: DW Innovation audio detection audit – the institutional reference that anchors the editorial frame for the entire 1B.3 cell.

In the toolkit's workflow

Cautious-detector pillar tool per Architectural Anchor 1; sits in 1B.3 audio deepfake forensics as an alternative on the strength of the DW audit ranking. The cell is structurally constrained: all three deployable audio detectors carry detector-as-weak-signal caveats, and the cell-level honest gap names lao-real-gap-audio, sinhala-tamil-asymmetric-gap, whatsapp-line-codec-untested, and detector-class-unreliable-cross-language.

This tool's verdict is one weak signal class per Anchor 1; never publish a binary claim from it alone. Counted as one detector signal class under Anchor 3 when paired with Hiya or TrueMedia.

Decision-tree references: T3 audio triage – Deepfake Total appears as the audit-leader pointer at T3.2 (speaker-identity check) feeding into T3.8 (audio detector run); the binding detector-as-weak-signal caveat applies at every publish-node check.

Override notes

Override notes

  • Stub required (stub-required-from-dw-audit-tool-042): the card ships as a stub because the publicly available documentation does not supply the documentation depth a full card needs. The stub framing is editorially deliberate and directs the reader to Hiya, TrueMedia, and DW audit for the operational and reference paths.

  • Vendor wrapping pair (c1-pair-vendor-docs-vs-dw-7of10): vendor materials are not summarised here in the absence of independent verification; the DW Innovation 7-of-10 finding is rendered in full in the Independent Accuracy admonition.

  • DW broken-across-SEA-languages verbatim (dw-audit-broken-sea-languages-verbatim): the DW conclusion is carried verbatim in the Limitations admonition rather than softened.

  • Weak-signal framing (e4-weak-signal-framing-binding): the workflow section above invokes the detector-as-weak-signal sentence per P3 sub-routine; never publish a binary claim from any audio detector alone, including Deepfake Total.

Sources