If a call just ended and you are still deciding whether to send money, skip everything below and read the red panel first. It works whether or not the voice was cloned, and whether or not you have a recording to check. The rest of this page covers what synthetic voices sound like, how to check a recording afterwards, and where to report it.
Before you check anything, before you research anything, before you decide whether the voice was real.
Cloned voices are convincing on timbre and accent — the things you consciously recognise. They are weaker on everything else.
| What to listen for | Why it happens |
|---|---|
| Breathing that is missing, or too regular | A person breathes irregularly and audibly, especially when upset. Synthetic speech either omits breath entirely or inserts it in a repeating pattern. |
| A line that is too clean | A distressed relative calling from a police station or a roadside has background noise. A generated clip often arrives with a silence cleaner than any real environment. |
| Emotion that does not track the words | Panic in the script but even, level delivery underneath — or distress that stays at exactly the same intensity for the whole call. |
| Trouble with interruption | Real conversation overlaps and recovers. A scripted or generated voice tends to talk over you, restart a sentence from the beginning, or leave an unnatural gap. |
| Odd handling of names and numbers | Proper nouns, account numbers and place names are where synthesis most often produces a stress pattern or pronunciation that is subtly wrong. |
| Deflection of specific questions | Bad line, no time, someone is listening, they will explain later. Anything that keeps you from asking something the script cannot answer. |
Ask an unexpected, specific question that could not be prepared and is not available online. Not a birthday, not a school name, not a pet — those are all scrapeable. Something like what you ate together last weekend, or what was on the table when you last met. A cloned voice is driven by an operator working from a script and public information, and it falls apart on that question almost immediately.
Better still, agree a family safe word in advance: a phrase a caller must produce before any urgent request is taken seriously. Agree it in person, keep it off every device and account, and use nothing that appears anywhere online. It costs one conversation and it removes the entire attack.
This is the part a detector helps with — after the fact, on a saved file, as one input among several.
You may have a recording if your phone records calls, if your organisation records inbound lines, or if the caller left a voicemail or sent a voice message rather than speaking live. Export it as MP3, WAV, M4A, OGG or FLAC — up to 10 MB — and run it through the AI Voice Detector. You get an AI-likelihood score and a human-versus-synthetic verdict in a few seconds, three checks a day, no signup.
If the message came through WhatsApp, it is an Opus file and needs converting first; the WhatsApp guide walks through it.
Telephone audio is the hardest case in this entire category. Calls are compressed aggressively and band-limited, which strips out much of the high-frequency detail detection depends on. Expect lower confidence than you would get on a clean recording, and read an uncertain result as genuinely uncertain rather than as a soft confirmation of what you already suspect.
To be explicit, because some marketing in this space implies otherwise: there is no real-time detection running on your phone line, from us or from anyone else at consumer scale. Anything sold as live protection is worth a hard look at what it actually does.
Report even if nothing was lost. Attempted-fraud reports are what let investigators connect a number or a script across many victims.
File at cybercrime.gov.in, the National Cyber Crime Reporting Portal.
Or call the cyber-fraud helpline on 1930.
If money has moved, call your bank first, then file. The golden-hour window is what makes a freeze possible.
Report to the FTC at reportfraud.ftc.gov.
For internet-enabled fraud, also file with the FBI's IC3 at ic3.gov.
Contact your bank or card issuer separately about any payment.
Report to Action Fraud at actionfraud.police.uk or call 0300 123 2040.
In Scotland, report to Police Scotland on 101.
Your bank's fraud line is the first call if money has left an account.
Voice cloning stopped being a specialist capability. Consumer tools can build a usable copy of a voice from a short sample, and the sample is rarely hard to get — a social media video, a voicemail greeting, a podcast appearance, a work webinar, or a recorded call placed for that purpose.
What makes it effective is not the audio quality. It is that hearing a familiar voice bypasses the part of your judgement that would normally slow you down. A text message claiming to be your daughter invites scepticism. Your daughter's voice, upset, does not — and the script is designed to keep you inside that first minute of reaction.
Common patterns worth recognising:
The defence in all four cases is identical and does not depend on any technology: end the call, look up the number yourself, call back, and confirm.
If the request arrived as a voice message rather than a live call — how to export it, convert it and check it.
Check a voice note →हिंदी में यही गाइड — the same guidance for Hindi-language calls and recordings.
हिंदी गाइड पढ़ें →Why phone audio is the hardest case, and how much weight a result deserves.
See our position →Three checks a day, free, no signup. Your audio is never stored.