The scenario
A finance officer joins a video call with someone who looks and sounds like a director, and authorises a transfer. Or a parent takes a call from a child in distress asking for money urgently. By the time the fraud is reported the account is emptied, the call has ended, and what remains is a recording, a screen capture, or nothing at all.
Why it is difficult
The window closes fast
Platform records, call metadata and the originating account are recoverable for days, not months. The technical question of whether the voice was synthesised is often answered long after the evidence that would corroborate it has expired.
The audio may be genuine
Not every impersonation is synthesised. An impersonator's real voice, a recut of the subject's genuine speech, or a hired actor all produce audio with no synthesis artefacts at all. A detector that only looks for vocoder traces will clear these — wrongly.
Call audio is the worst case for detection
Telephony and VoIP band-limit speech, apply aggressive noise suppression, and encode at low bitrates. Every one of those strips the evidence that synthetic-speech detection depends on, and they do it to genuine calls too.
How Exiphore approaches it
Examine what survives the channel
Acoustic measurements are weighted for the channel they arrived through. Where telephony has destroyed the high-frequency band, those measures are discounted rather than allowed to read compression as authenticity — and the report says so explicitly.
Separate the voice question from the video question
On a video call exhibit, the audio and the face are assessed independently. Real audio over a substituted face is a common construction, and treating a genuine voice as evidence that the video is genuine is exactly how these cases get cleared in error.
Preserve first, analyse second
Every examination produces an action plan whose first steps are preservation requests and source-device seizure. Those are generated by rule and appear regardless of what the technical finding turns out to be, because they decay fastest.
Link the campaign
Fraud of this kind is rarely a single incident. Fingerprints from each exhibit are indexed, so a recurring synthetic voice or a shared production toolchain surfaces across cases as one operation rather than several.
What this will not settle
- The platform does not perform speaker identification. It can indicate whether speech shows signs of synthesis; it cannot tell you whose voice it is. That requires an enrolled reference sample and a separate examination.
- A clean result on a compressed call recording is weak evidence. The channel removes what the analysis needs, and the report states the resulting ceiling.
Bring us an exhibit you already know the answer to
The most useful demonstration is on your own material, from a closed case where the ground truth is settled.
Request a demoRelated use cases
Identity and video-KYC abuse
Synthetic faces defeat liveness checks and open mule accounts at scale. Examine onboarding captures and link recurring synthetic identities.
Evidentiary review and disclosure
Before an exhibit is tendered, establish what can and cannot be said about it — measurements, limitations, custody and signed reasoning as one sealed record.