BabelFake
BabelFake: A Multilingual Audio-Visual DeepFake Benchmark
A multilingual audio-visual DeepFake benchmark with 399k clips from 496 consenting participants across 5 languages, using modern manipulation pipelines.
How to run
$ npm i -g lettokenburn-cli$ ltb bench info babelfake
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
6 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| SpeechForensics | 79.8 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| RAVEn-AV | 72.0 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| RAVEn-V | 70.3 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| AVH-Align sup | 67.1 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| RAVEn-A | 60.1 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
| AVH-Align | 18.5 | N/A | Independent | medium | Oct 5, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsAbout
BabelFake comprises 399k clips (1,323 hours) from 496 individuals speaking English, German, Italian, French, and Spanish. It uses a modular pipeline pairing 11 video manipulation methods with 4 voice cloning engines.
The benchmark evaluates state-of-the-art detectors, revealing performance drops when visual manipulations are paired with authentic audio. It highlights language sensitivity in detector architectures and misalignment between human perceived realism and machine detection difficulty.