Authentic video, authentic voice
- Video: appears authentic
- Audio: appears authentic
Neither track shows the signals we look for.
Nothing to action. Keep the report on file.
Deepfakes rarely fake everything. Now you can tell what they faked.
SyncLens checks the face and the voice separately, then tells you whether they agree. Upload a video, an audio file, or both. Find out which track was faked, if either was.
Assembled
Modern fakes are built from parts. A synthetic face over a real voice, or a real face reading a script in a voice that was cloned.
Split
The face and the voice can be manipulated on their own. One clip can be genuine on one track and generated on the other.
Unactionable
A yes-or-no answer tells you something is off. It does not tell you what changed, so it does not tell you what to do next.
Four outcomes, four responses
SyncLens returns one of four results, each with its own confidence band. What you do next depends on which one it is.
Each result carries a shape as well as a label, so it reads without colour. The first marker is the video track, the second is the audio track. A circle means nothing stood out. A triangle means the track appears synthesized.
Synthetic video, synthetic voice
- Video: appears synthesized
- Audio: appears synthesized
The face and the voice both look generated.
Treat the whole clip as synthetic.
Synthetic video, authentic voice
- Video: appears synthesized
- Audio: appears authentic
A swapped or synthetic face over a real recording of someone speaking.
The words may be genuine. The person on screen may not be.
Authentic video, synthetic voice
- Video: appears authentic
- Audio: appears synthesized
A real recording with a synthetic voice laid over it.
The footage is genuine. The script is not.
The pipeline, start to finish
Face detection and alignment
Every frame with a face is found, cropped, and squared up so the later stages compare like with like.
Three visual reads
Identity consistency from frame to frame. Pixel-level texture. Motion and blink patterns over time.
Two acoustic reads
Fine spectral detail in the voice. Rhythm and intonation across the whole clip.
Synchronisation check
Cross-attention lines up lip movement against sound. Genuine speech stays coherent. Two tracks faked apart come loose, and that shows on the timeline.
Classification
The evidence is combined into one of the four results, with a calibrated confidence band.
Every model in this pipeline is frozen and versioned. A report produced today can be reproduced tomorrow.
We show the work
Every report includes the material behind the result, not just the result.
- Tagged frames with the detected face boxed.
- Heatmaps over the regions that drew the model's attention.
- A per-segment timeline of audio-visual sync.
- The analysed mel-spectrogram, with silence greyed out.
- A provenance panel: model version, input file hash, processing time.
Three kinds of evidence in every report
Frame evidence
Every detected face, boxed, with a per-frame suspicion score and a Grad-CAM heatmap you can turn on.
Sync timeline
Audio-visual correspondence across the whole clip. Where two tracks were faked apart, it visibly breaks up.
Provenance
Model version, input file hash, media duration, processing time. Enough to reproduce the report later.
Built for the people who act on the answer
Five ways teams put a four-class report to work. Each has its own page.
Fits the tools you already use
Send media from your stack to the API when a recording lands, and get a report back before anyone needs it.
Applicant tracking
Video interview platforms
Moderation queues
Case management
Cloud storage
Webhooks
REST API
CI pipelines
One call. A four-class answer.
Request
curl -s https://api.synclens.io/v1/jobs \
-H "Authorization: Bearer $SYNCLENS_KEY" \
-F "media=@interview.mp4"Response
{
"job_id": "job_9f2c...",
"status": "completed",
"result": {
"class": "FakeVideo-RealAudio",
"confidence": { "band": "moderate", "p": 0.71 },
"probabilities": {
"RealVideo-RealAudio": 0.06,
"FakeVideo-FakeAudio": 0.18,
"FakeVideo-RealAudio": 0.71,
"RealVideo-FakeAudio": 0.05
},
"segments": [
{ "start_ms": 0, "end_ms": 5000, "class": "RealVideo-RealAudio", "sync": 0.92 },
{ "start_ms": 5000, "end_ms": 10000, "class": "FakeVideo-RealAudio", "sync": 0.34 }
],
"model_version": "v1.0",
"input_sha256": "b1946ac9..."
}
}What this does not do
Detection is probabilistic. Performance drops on heavy compression, on very short clips, and on generators the model has not seen before. A result is forensic evidence, not proof, and it should be read alongside other context. When the model is unsure, the report says so before it says anything else.
Read the known limitationsPay for the minutes you analyse
Billed by the minute of media analysed, not per file. Early-access pricing while we measure real cost.
Free
$0
10 minutes a month. Web only. Watermarked report.
Pro
$29
120 minutes. API access. PDF export. 30-day retention.
Team
$149
750 minutes. Batch upload. Webhooks. Five seats.
Questions people actually ask
How accurate is it?
We do not publish an accuracy number. A single figure hides how much performance changes from one clip to the next, and it goes stale the moment a new generator appears. Every result comes with a calibrated confidence band instead, and low confidence is stated plainly.
What happens to my uploads?
Raw files are deleted automatically after seven days on the free tier, and sooner on request. Only the derived report is kept, for your plan's retention window. Files are encrypted at rest.
Can I use this in court?
The forensics tier is built for reproducibility: versioned models, hashed inputs, exportable reports. Whether a specific report is admissible is a question for the court and your counsel.
What if it is wrong?
It will sometimes be wrong, in both directions. That is why the output is evidence with a confidence band rather than a verdict. Read the per-segment view and the frame evidence, not only the headline result.
How teams use it
Four workflows the four-class report drops into.
Pre-offer screening
Run recorded final-round interviews through before an offer goes out.
Upload-time moderation
Check user uploads for synthetic media before they reach an audience.
Casework
Attach a reproducible report to an investigation, with the model version pinned.
In your product
One API call inside your own pipeline. A four-class response, per segment.
