How much audio does someone need to clone my voice?

Seconds. Modern zero-shot cloning works from a clip shorter than most Instagram Reels, and needs no training run. Any assumption that you are safe because you have only posted a little audio is out of date.

Can I tell a cloned voice from a real one on a phone call?

Not reliably, and not under stress. Phone compression removes the artefacts you would listen for, and the scam is specifically designed to prevent careful listening. This is why the defences that work are procedural — a callback on a saved number, an agreed verification phrase — rather than perceptual.

Does watermarking stop misuse?

It does not prevent it; it makes it traceable and provable after the fact. That matters for takedowns, for disputes and for investigations, which is why we keep the watermark rather than offering to strip it.

Safety

AI voice scams in India, and what actually protects you

We build voice cloning, which makes this awkward to write and also means we know exactly how the attack works. Here is the honest version, including what our own industry should be doing and largely is not.

9 min read · 25 August 2026

The technology that lets a creator narrate a video in their own voice is the same technology that lets someone call your mother in your voice. There is no version of this product that has one capability and not the other, and pretending otherwise would be dishonest.

So: how the fraud actually works, why India is a particularly good target, and what genuinely helps.

The attack, concretely

Source audio is collected. Instagram Reels, YouTube, a WhatsApp voice note forwarded onward, a wedding video. Modern zero-shot cloning needs seconds, not minutes, and public posting supplies it for free.
A clone is made. On any of a dozen tools, several of which ask nothing about consent.
A pretext is chosen. Almost always urgency plus authority plus isolation: an accident, an arrest, a medical emergency, a deadline that expires in minutes.
The call is placed, usually from an unknown number, usually with poor audio quality — which is a feature, because phone-line compression conceals exactly the artefacts that would give a clone away.
Payment is requested by a route that settles instantly and reverses with difficulty.

Why this works especially well in India

Three structural reasons, none of which are about gullibility:

UPI settles instantly. A transfer made under pressure is gone before the pressure lifts. Card fraud has chargebacks; a UPI payment made voluntarily has a much harder recovery path.
WhatsApp carries voice by default. Voice notes are the normal register of family communication here in a way they are not in many markets, so an unexpected voice message is unremarkable.
Family finance is genuinely urgent-shaped. Medical emergencies requiring immediate payment are a real and common experience, which means the pretext is not implausible. The scam borrows credibility from a true feature of life here.

What actually protects a family

Agree a family verification phrase, now, in person. A word or question with an answer no public post reveals. Not a birthday, not a pet's name — those are on your timeline. Then use it whenever money is discussed by voice, without exception and without embarrassment.
Make a hard rule: hang up and call back on the saved number. Not the number that called. Not a number the caller gives you. The one already in your phone. A clone cannot answer a call you place.
Treat urgency itself as the signal. Real emergencies survive a two-minute delay. Manufactured ones cannot afford one, which is why every version of this scam insists there is no time.
Never transfer under time pressure to an unfamiliar UPI ID or account. If it is genuinely your family, they will still be your family in ten minutes.
Tell the people most likely to be targeted. Parents and grandparents, specifically, and tell them the mechanism rather than just "be careful" — people defend better against a threat they understand than against a warning.

Reducing your exposure, without deleting your life

Advice to stop posting your voice is useless to anyone who publishes for a living, and mostly useless to everyone else. More realistic:

Lock down old accounts you no longer use but that still host audio of you.
Be more careful with voice notes in large groups than with public content — a forwarded voice note has a smaller, more targeted blast radius and reveals a more intimate register.
Understand that being a public creator means your voice is already clonable, and put your effort into the verification procedures above rather than into containment that has already failed.

What platforms should be doing

This is where we have a position, and where the industry's record is poor:

Consent should gate cloning, verified server-side. A checkbox that the interface can skip, or that an API call can omit, is decoration.
Submitted reference clips should be screened against known voices, with matches held for human review rather than auto-approved. This is the step that catches public-figure cloning at submission.
Output should stay watermarked. Inaudible watermarking is the one provenance signal that survives re-encoding and re-upload, and stripping it — which some tools advertise as a premium feature — is actively harmful.
There should be a monitored abuse route with a real human behind it, published where a victim can find it without an account.

We do all four, and we would rather compete with platforms that also do. What we cannot do is stop someone using a tool with none of these controls, which is why the family-level procedures above matter more than any vendor's policy page.

If it has already happened

Report the financial fraud immediately on the national cybercrime portal at cybercrime.gov.in, and to your bank. Speed genuinely affects recovery odds in the first hours.
Call 1930, the national cyber-fraud helpline, for a payment made in the last few hours.
Preserve everything — the audio, the number, the transaction reference, the timestamps — before deleting anything.
Report the misused voice to the platform that generated it, if you can identify it. Under Indian intermediary rules, platforms must publish a grievance contact and respond within defined timelines; ours is in the footer of every page on this site.

If your own voice has been cloned without your consent and used on Mayantar, that is a report we prioritise over everything else in the queue. The address is on the safety page.