Home › Voice Detection › For Fraud Teams

Cloned voices and authorisation fraud.

For a fraud or security function the useful question is not “can we detect deepfakes” but “where does our process still treat a voice as proof of identity?” Voice has stopped being an authentication factor, and the controls that follow from accepting that are procedural. Detection has a real but narrow role, mostly after the fact.

Check an audio clip free The controls that work ↓
First check free, no signup Audio never stored
Known issue with the current detector — please read before acting on a result. The engine running in production over-flags genuine recordings: real human speech, especially compressed audio such as phone calls and messaging-app voice notes, can come back as high-confidence “AI-generated”. A replacement engine is in acceptance testing. Until it is live, treat a synthetic verdict on a recording you believe is real as unproven, and rely on verifying the speaker directly — calling back on a number you look up yourself settles it in a way no detector currently can.
Premise

Voice is no longer an authentication factor.

Everything sensible follows from accepting this, and nothing sensible follows from resisting it.

A recognisable voice used to be reasonable evidence that a known person was on the line. It is not any more, and no detection capability restores it — the arms race structurally favours generation, because synthesis improves faster than detection corpora are rebuilt.

So the design question is not how to spot cloned voices. It is where your processes currently allow a voice, on its own, to authorise something. Common places that survive longer than they should:

  • Callback verification where the agent dials a number supplied during the same interaction
  • Payment release on verbal authorisation from a recognised internal voice
  • Password and MFA resets where an agent is satisfied by knowing the person
  • Voice biometrics used as a sole factor rather than as one signal among several
  • Exception handling where an urgent, senior-sounding caller can bypass the normal route
Voice biometrics deserve specific attention. Speaker-verification systems answer “does this match the enrolled voiceprint”, which is a different question from “is this a live human”. Anti-spoofing is a separate capability that has to be deliberately present and tested. If you deploy voice biometrics, ask your vendor directly what their spoofing countermeasure is and how it was evaluated.
Controls

What actually stops this fraud.

None of these depend on detecting anything.

  1. Out-of-band callback, on your own number. The single highest-value control. Never a number supplied during the interaction, never a transfer offered by the caller. This defeats essentially every inbound cloned-voice attack.
  2. Dual authorisation above a threshold, through separate channels. A second approver contacted independently, not the person the first caller nominated.
  3. No exceptions for urgency or seniority. Urgency and authority pressure are the attack. A process that bends for a senior voice in a hurry has no floor, and staff need explicit permission to hold the line.
  4. Fixed payment-change procedure. Bank-detail changes verified through a pre-established channel with a mandatory delay, never on a single call.
  5. Train on the pattern, not the technology. Staff do not need to identify synthesis by ear. They need to recognise the shape — urgency, secrecy, authority, an unusual route — and to know that following procedure will never be held against them.
The tell that survives everything: a request to bypass the normal process. Whether the voice is real or cloned is almost irrelevant if the process holds.
Where detection fits

The narrow but real role of analysis.

Detection earns its place after the event rather than during it.

Incident review

Where a call was recorded, checking the audio adds a line to the incident record and helps characterise what happened. That matters for internal reporting, for insurance and for law enforcement referral — as supporting context, explicitly not as a determination.

Pattern building

Across a set of related incidents, consistent findings help establish whether you are seeing one campaign or several. Individually weak signals can be collectively informative.

Training material

Genuine examples from your own environment are considerably more effective in awareness training than generic ones, and a check helps you label them correctly before you use them.

What it cannot do

  • Run on a live call. Detection needs a recorded file. There is no real-time consumer capability here, from us or anyone else.
  • Deliver reliable results on telephone audio. Call codecs are band-limited and aggressive, and they strip the detail detection depends on. This is the hardest audio class in the category — set expectations accordingly.
  • Name the generator. Our detector returns human-versus-synthetic, not attribution.
  • Support a determination against a person. Not forensic-grade, and never a sole basis for a financial, legal or disciplinary decision.
Operational note: retrospective analysis can be automated — POST /v2/voice/detect takes an audio file and is included with the Broadcast voice plan, so a recorded-call queue can be scored programmatically rather than by hand — one request per file; there is no bulk or asynchronous submission. The API page has the endpoint reference. There is still no live-stream detection, so this remains an after-the-fact review, not call screening.
Not forensic-grade. A TextSight voice result is a triage signal, not proof, and it must never be the sole basis for a disciplinary, employment, financial or legal decision. Our full position on what a result is worth is here.
FAQ

Fraud team questions.

Can we screen inbound calls for AI voices in real time?
No. Detection runs on recorded audio after the fact. There is no real-time capability here from us, and claims of live protection are worth interrogating closely. Out-of-band callback on a number you control is the control that actually works during a call.
Is voice biometrics still safe as an authentication factor?
Speaker verification answers whether audio matches an enrolled voiceprint, which is a different question from whether the speaker is a live human. Anti-spoofing is a separate capability that must be deliberately present and independently evaluated. Ask your vendor what their countermeasure is and how it was tested, and do not rely on voice as a sole factor.
How reliable is detection on recorded phone calls?
Less reliable than on clean audio, and you should plan for that. Telephony codecs are band-limited and heavily compressed, removing much of the fine spectral detail detection relies on. Treat uncertain results as uncertain and use a known-human call through the same system as a control baseline.
Can we use a result in a law enforcement referral?
As supporting context, described accurately as an automated assessment that is not forensic analysis. It should not be characterised as a determination. Where a case justifies it, a qualified forensic audio examiner produces something that can be defended under challenge.
Can we score recorded calls through an API?
Yes, for recorded audio: POST /v2/voice/detect accepts an audio file and is included from the Amplify voice plan up, so a queue of recorded calls can be scored programmatically. It is one file per call, synchronous, with no bulk or asynchronous submission today. There is no live-stream endpoint, so this is after-the-fact review rather than call screening.
What single control should we implement first?
Out-of-band callback on a number your organisation controls, with an absolute rule that urgency and seniority never justify an exception. It is cheap, it needs no technology, and it defeats essentially every inbound cloned-voice attack regardless of how good the synthesis gets.
Related

More voice detection guides.

How accurate is it, really?

Why we publish no single accuracy percentage, what false positives look like, and how to weigh a result.

See our position →

Voice detection API

What the public API covers today, what it does not, and how to get in touch about volume.

See the API status →

Further reading

Reviewing a recorded call? Add it to the file.

Your first check is free with no signup — 3 a day with a free account. Your audio is never stored.

Check an audio clip All voice guides
Retrospective only · no live-call detection · phone audio is the hardest case