Free Deepfake Detector | Video & Voice Clones | LegalAudit

Deepfake video and voice-clone scams jumped 1,740% between 2022 and 2025 (Sumsub Identity Fraud Report 2025). The technology is now cheap enough that ordinary romance, recruitment, CEO-fraud and family-emergency scams routinely arrive as 7-second video clips or 30-second voice messages — and that's where most people freeze. The Deepfake Detector is built for that 90-second window: drop the file in, get a verdict you can act on.

Under the hood we run three independent classifiers (frame-level temporal inconsistency, lip-sync vs phoneme alignment, voice spectral fingerprinting against known synthesis stacks like ElevenLabs, Resemble AI, OpenVoice and PlayHT), then cross-validate with C2PA / SynthID provenance signals when present. The verdict is conservative — we flag GREEN, AMBER, or RED with an explicit confidence percentage and the specific frames or audio segments that triggered the call, so a lawyer or compliance officer can sanity-check the reasoning.

How it works

Upload the video or voice note

Drag-and-drop the file into Mythos. We accept MP4, MOV, WebM, MP3, M4A, OGG, WAV. Up to 200 MB on free plan. The file is processed in-memory; the original is hashed (SHA-256) and discarded unless you generate the dossier.

Provide one-line context

Tell Mythos what the file purports to show — a colleague, a relative, a CEO message, an investment pitch. Context steers the temporal-coherence model toward the relevant face/voice region.

Read the per-frame / per-segment verdict

Mythos returns GREEN / AMBER / RED with a confidence %, plus a timeline of the suspicious frames (video) or audio segments (voice) and a plain-language explanation of which classifier triggered the call.

Optional: generate the dossier

If the analysis matters for a complaint, civil claim, or workplace incident, generate the forensic dossier — hash-locked exhibit chain, methodology citation (NIST SP 800-86 + ISO/IEC 27037), Swiss case-number format.

What we detect

GAN-based face swap (FaceSwap, SimSwap, Roop variants)

Diffusion-based face reenactment (latest 2025 models)

Lip-sync / phoneme alignment mismatch

Eye-blink rate and gaze inconsistency

Skin texture and pore-density irregularities under zoom

Voice clone signatures (ElevenLabs, Resemble, PlayHT, OpenVoice)

Spectral discontinuities across phoneme boundaries

Compression artefacts from generative pipelines

Background bleed / breath gap inconsistency in voice notes

C2PA / SynthID provenance verification when present

Frequently asked questions

How accurate is the deepfake detector?

On our 2026-Q1 internal benchmark covering 1,200 manipulated and 1,200 genuine samples across video and voice, the ensemble verdict reached 96.3% precision and 91.7% recall at the conservative threshold (RED = high confidence). False positives are heavily penalized; we'd rather flag AMBER than wrongly accuse. Adversarial cases (custom-trained models, very short clips < 3 seconds, heavily compressed audio) remain harder — those typically come back as AMBER with explicit caveats.

Does the detector work for very short clips?

Yes, but with caveats. Under 3 seconds for video and under 5 seconds for voice we drop confidence by ~15 percentage points and flag the limitation explicitly in the verdict. For the most reliable verdict, give us the longest unedited segment you have — even bad lighting or compressed clips help, as long as they aren't trimmed to a single sentence.

Is the analysis admissible as evidence?

The forensic dossier is structured to be admissible as expert technical opinion under Swiss CPP art. 184 and EU equivalents (Daubert/Frye in the US). It cites NIST SP 800-86 (computer forensic guidance), ISO/IEC 27037 (digital evidence identification), the specific models and thresholds used, and includes SHA-256 hashes of the inputs and outputs. For higher-stakes cases (criminal trial, regulator investigation) the CHF 490 Expert Review adds a human forensic analyst signature.