Can AI clone someone's voice from a short clip?
YesYes. A few seconds to a few minutes of someone's recorded speech is enough for widely available tools to produce a convincing copy of their voice.
First shown: Jan 5, 2023. Microsoft researchers' VALL-E produced speech in the voice of an unseen speaker from a 3-second recording. (VALL-E paper (arXiv))
What it means
Voice cloning means software learns how one person sounds (pitch, accent, rhythm) from a recording and then says anything you type in that voice. In January 2023 Microsoft's VALL-E research system did this from a 3-second sample, without having been trained on that speaker.
Within weeks it left the lab. ElevenLabs opened a public beta in January 2023 that cloned voices from short snippets, and had more than a million registered users by June 2023. In February 2023 a journalist used its free service, with about five minutes of his own speech, to get past his bank's voice check.
The limits are about quality, not possibility. Short clips give a rougher copy than long ones, unusual accents are harder, and a live back-and-forth is harder to fake than a pre-written message. But a frightened parent hearing a crying child on a bad line is not listening carefully, which is why scammers use it.
Safeguards are thin. In March 2025 Consumer Reports found that four of six popular voice-cloning services failed to take basic steps to stop people cloning a voice without its owner's consent.
The evidence ladder
Every capability on this site is graded on the same five steps, from a single demonstration to real-world consequences.
| Step | Status | When | Evidence |
|---|---|---|---|
| Demonstrated shown at least once | yes | Jan 5, 2023 | Microsoft's VALL-E paper showed speech in an unseen speaker's voice from a 3-second recording; the model was not released. (VALL-E paper (arXiv)) |
| Reproduced independently repeated | yes | Jan 31, 2023 | ElevenLabs' public beta (January 2023) cloned voices from a few short snippets, and abuse with celebrity voices was reported by Jan 31, 2023. (Wikipedia: ElevenLabs) |
| Accessible available to ordinary people, or criminals | yes | Mar 10, 2025 | Cheap or free consumer apps offer cloning; Consumer Reports found four of six leading services (Descript, ElevenLabs, Lovo, PlayHT, Resemble AI, Speechify) lacked basic consent safeguards. (Consumer Reports) |
| Reliable works most of the time for a defined task | yes | Feb 23, 2023 | A cloned voice passed Lloyds Bank's voice ID in 2023, though only after several failed attempts and with the account holder's date of birth; commercial cloning has run at scale since. (Vice) |
| Consequences real-world use or harm documented | yes | Oct 6, 2026 | Fake family-emergency calls (2023), the AI Biden robocall that drew a $6M FCC fine (2024), and Dutch bank-helpdesk fraud opened by AI-generated calls that cost about 170 victims some EUR 380,000 (2026). (NL Times) |
From the record
- An Arizona mother got a fake kidnapping call in a voice indistinguishable from her teenage daughter's; she later testified to the US Senate. The case file → (US Senate testimony)
- Robocalls with an AI-cloned voice of President Biden told New Hampshire Democrats to skip the primary; the FCC fined the consultant behind them $6 million. The case file → (NPR)
- Dutch police arrested two men over bank-helpdesk fraud that opened with AI-generated calls; about 170 victims lost about EUR 380,000 in three weeks. The case file → (NL Times)
- Nigerian police allege a 23-year-old used AI voice cloning and deepfakes to pose as Elon Musk in a crypto investment scam. The case file → (Blueprint (Nigeria))
What it means for you
A voice on the phone is no longer proof of who is speaking, even someone you know well. Any public video, voicemail or social media clip can be enough raw material.
What to do
- Agree a family code wordPick a word or question only your family knows and ask for it on any urgent call about money or danger.
- Hang up and call backIf a call asks for money urgently, end it and ring the person on the number you already have.
- Distrust urgency and secrecyPressure to act now and tell no one is the scam's engine, whatever the voice sounds like.
- Drop voice-only bank loginIf your bank lets your voice act as a password, ask for a stronger option such as an app or PIN.
Questions people ask
How much audio does AI need to clone a voice?
Research systems have worked from 3 seconds; consumer tools ask for a short clip, and a few minutes of audio gives a closer match.
Can I tell a cloned voice from a real one?
Not reliably by ear, especially on a phone line and under stress. Verify by calling back or asking a code word instead.
Do voice-cloning companies stop misuse?
Some do more than others. Consumer Reports found in March 2025 that four of six popular services lacked basic steps to confirm the speaker consented.
Is using a cloned voice to trick people illegal?
Using it to defraud someone is fraud. In the US, AI-voiced robocalls were ruled illegal under the Telephone Consumer Protection Act in 2024.
Related
- AI voice clone scam call: what to do right now →
- How to set up an AI chatbot for an older parent, safely →
- AI scam texts: how to spot them and what to do (2026) →
Sources
- VALL-E paper (arXiv, Jan 2023)
- Wikipedia: ElevenLabs
- Consumer Reports on voice-cloning safeguards (Mar 2025)
- Vice: How I broke into a bank account with an AI-generated voice
- DeStefano Senate testimony (June 2023)
- NPR on the Biden robocall fine
- NL Times on Dutch AI bank-helpdesk fraud
Spot an error or something out of date? Tell us; corrections are made and logged. Written with AI assistance from public sources; every fact checked and approved by Colin Balfe. General information, not legal, medical or financial advice.