The problem has moved
Voice biometrics were built for a world where imitating someone took talent. A few seconds of a customer’s audio now buys a working clone, and knowledge questions are answered from data that has already been breached. At the same time, more of the front line is handled by AI agents that cannot judge a voice by ear.
Regulators have noticed too: in 2024 the US FCC ruled that AI-generated voices in robocalls fall under existing robocall rules, and the EU AI Act introduces transparency obligations for synthetic content. Contact centers need a way to know when the voice on the line is generated.
Where attacks happen in the call flow
- IVR reconnaissance. Attackers map which accounts respond long before they reach an agent.
- Authentication. A recording or clone of the right customer clears a voiceprint gate.
- Human agents. Asked to judge authenticity by ear, under handle-time pressure, on compressed telephony audio.
- AI agents. Act on the transcript, which looks identical for a real or cloned caller.
Score before the queue — and throughout
Tap the media stream — SIP/RTP or your CCaaS leg — and send it to the Vosco API. The first verdict lands within 3 sec of the caller speaking — before the call reaches a queue — and a fresh one follows every 3 sec after. Scoring continues for the whole call, so a mid-call switch to a recording or a clone is caught too. The same pattern for programmable telephony is described in our Twilio guide.
Give agents signals, not a verdict to argue with
Put the per-signal breakdown on the agent desktop — synthetic, replay, speaker and liveness — rather than a single opaque score. Agents and fraud teams can see why a call was flagged, and policies can treat “replay suspected” differently from “synthetic speech detected”.
Keep your voice biometrics
Voice trust does not replace existing biometrics. Biometrics answer which speaker; Vosco also answers whether a live human is speaking at all. They stack — see verification vs. identification and replay vs. synthetic attacks.
Cloud or on-prem
Teams that can use the cloud start on the usage-based API. Banks, telcos and regulated operators where call recordings cannot leave the estate run Vosco Precise on-prem on an annual licence, at 1,000+ concurrent streams with under 10 ms GPU inference.
Frequently asked questions
How do contact centers detect deepfake callers?
By scoring the call audio itself for synthetic speech, replay and speaker identity in real time, from the first seconds of the call and throughout the conversation.
Does voice trust replace voice biometrics?
No. Voice biometrics identify which speaker it is; voice trust also checks whether a live human is speaking. They work together.
Can Vosco run on-premise in a contact center?
Yes. Vosco Precise runs on-prem on an annual licence, supporting 1,000+ concurrent streams for environments where recordings cannot leave the estate.
Score every call before the queue.
Start on the cloud API, or talk to us about running Vosco Precise on-prem.