Home · Voice cloning

What is voice cloning? How it works, and how to catch it.

Updated November 2026 · Reviewed against FTC, FBI and FCC sources dated 2023 to 2026

Voice cloning is the use of AI to copy a specific person's voice from a recording, so that software can say new words in that voice. A cloned voice can read any script, in real time or as a saved file. The same technology powers audiobooks and accessibility tools, and it is now the engine behind family emergency scams, CEO fraud and AI vishing calls.

Definition

What is voice cloning?

Voice cloning, also called voice replication or a voice deepfake, is a form of synthetic speech that imitates one particular person rather than a generic narrator. Ordinary text-to-speech gives you a stock voice. A voice clone gives you someone's voice: their timbre, their pitch range, their accent and some of their rhythm.

Cloning used to need hours of studio recordings. Modern models need far less. Microsoft's VALL-E research paper showed speech synthesised from a three-second enrolment recording of an unseen speaker (Wang et al., January 2023), and consumer services now sell cloning from a short upload. The FTC's March 2023 alert put the consequence plainly: a scammer can clone a loved one's voice from a short clip posted online (FTC consumer alert).

In plain terms: if your voice is on a voicemail greeting, a podcast, a conference talk or a social video, there is enough of it to clone.

Mechanics

How does AI voice cloning work?

Every cloning system, commercial or open source, runs the same four stages. The details differ; the shape does not.

  1. 01
    Collect a sample. A recording of the target speaking. Clean, single-speaker audio works best, which is why podcasts, interviews and voicemail greetings are favourite sources.
  2. 02
    Encode the voice. A neural network converts the sample into a compact numerical description of the speaker, often called a speaker embedding. It captures what makes the voice recognisable, separate from the words spoken.
  3. 03
    Generate new speech. A speech model takes new text, or a second person's live speech, and produces audio conditioned on that embedding. The output says words the target never said, in something close to their voice.
  4. 04
    Render the waveform. A vocoder turns the model's output into an audio file or a live stream. This last step is where many of the acoustic traces that detectors look for are introduced.

Quality rises with more and cleaner input. A short sample gives a passable likeness over a phone line, where compression hides much of the difference. A longer, clean sample gives a clone that can pass on a good speaker.

Comparison

What are the types of voice cloning?

Three approaches cover nearly everything in use today. They matter because each leaves a different kind of evidence and enables a different kind of attack.

TypeHow it worksInput neededTypical useWhat it means for fraud
Zero-shot (instant) cloningA pre-trained model imitates a new voice from a short sample, with no retrainingSeconds to a minute of speechConsumer apps, quick voiceoversThe main scam engine. A clip from social media is enough to fake a relative on a phone call.
Fine-tuned (professional) cloningThe model is further trained on the target's recordingsTens of minutes to hoursAudiobooks, dubbing, licensed brand voicesHigher fidelity. Used against public figures with lots of available audio, such as executives and politicians.
Voice conversion (speech-to-speech)A person speaks live and the system re-voices them as the target in near real timeA sample of the target plus a live speakerVoice changers, live dubbing, gamingEnables live, interactive scam calls. The scammer can answer questions instead of playing a script.
Legitimate uses

What is voice cloning used for legitimately?

Most cloning is not crime. It restores speech for people who are losing their voice to illness, it dubs film and training video into other languages in the original speaker's voice, it lets authors narrate audiobooks without weeks in a booth, and it lets brands license a consistent voice. Reputable services require consent from the person being cloned and add safeguards such as voice verification.

The problem is that the safeguards only bind honest users. Open-source models run on a laptop with no consent check, and a consenting upload can be faked. That is why detection and verification now sit on the receiving end, with the person who hears the voice.

Voice cloning scams

How do voice cloning scams work?

A voice cloning scam is a phone call, voice note or voicemail in which a cloned voice of someone you trust asks for money, a code or access. It is vishing with the weak point removed: the voice now sounds right. These are the patterns that account for most reported cases.

Family emergency

A grandchild or child "calls" after a crash or an arrest and begs for bail or a lawyer's fee, often handing over to a fake officer or attorney. The FTC warned of exactly this in March 2023 (FTC). See the grandparent scam.

Executive (CEO) fraud

A cloned executive tells finance to wire funds for a confidential deal. In 2024 an Arup employee sent about US$25.6 million after a video call in which the other participants were deepfakes of real colleagues (CNN, May 2024). See CEO fraud.

Official impersonation

In May 2025 the FBI warned that criminals were sending AI-generated voice messages posing as senior US officials to build trust, then steering targets to links that took over their accounts (FBI PSA I-051525-PSA).

Bank and account takeover

A cloned voice is used against a bank's phone line or help desk to reset a password, change a payee or pass a "voice is my password" check. The same clone can call the customer, posing as the bank's fraud team.

Virtual kidnapping

The caller plays a cloned voice screaming or crying, claims to be holding the person, and demands a fast ransom. The person is safe; the panic is the payload.

Political and robocall deception

In January 2024 a robocall using an AI imitation of President Biden told New Hampshire voters not to vote in the primary. The FCC responded by ruling AI-generated voices "artificial" under the Telephone Consumer Protection Act (FCC, February 2024).

The scale is now measured. The FBI's 2025 Internet Crime Report included a section on artificial intelligence for the first time: 22,364 complaints and nearly $893 million in losses tied to AI-enabled schemes, including voice clones (FBI, April 2026). For the full set of scripts, see AI voice scams and real vishing examples.

Red flags

How can you tell if a voice is cloned?

By ear, often you cannot. The FBI's own guidance says AI-generated content "is often difficult to identify" (FBI, May 2025). Listen for these anyway, and weigh the situation more than the sound.

  • 01Emotion that does not fit. Panic in the words, but a flat or oddly even delivery, or crying that loops.
  • 02Breathing and pauses in the wrong places, or none at all between long sentences.
  • 03A room that never changes. No background shift when the caller says they moved, no echo where there should be one.
  • 04Mispronounced names, places or family nicknames the real person would say without thinking.
  • 05Delays before each answer on a live call, as if the reply is being typed or processed.
  • 06A refusal to switch to video, or to let you call back on their usual number.
  • 07The script itself: urgency, secrecy and an unusual payment. That combination matters more than any audio tell.

None of these is proof. Phone compression, a bad signal or real distress can produce every one. Treat them as reasons to verify, not as a verdict.

Detection

Can AI detect a cloned voice?

Yes, from a recording. A cloned voice is synthetic speech, and synthetic speech carries statistical traces that a trained model can pick up even when a listener cannot: patterns in the spectrum, the vocoder's fingerprint, prosody that is too regular. An AI voice detector scores those traces.

Here is what AI Voice Detector does with a saved clip such as a voicemail, a voice note or a call recording, and what it does not do.

It doesIt does not
Returns one of four verdicts: AI generated, likely AI, likely human or humanScreen a live call while it is happening
Gives an AI probability score for the whole clipTell you whose voice was cloned, or which tool made it
Scores each segment, so a real voice with a cloned sentence spliced in stands outProve a recording is authentic. A human verdict lowers the risk; it does not clear the request
Checks a single clip free, up to 10 MB and 60 seconds, without an accountReplace the callback to a number you already hold

For a step-by-step method that combines listening, context and the detector, see how to detect AI voices. Teams screening recorded calls at volume can use the API or read how fraud teams use it.

Protection

How do you protect yourself from voice cloning?

You cannot keep your voice entirely private. You can make a clone useless.

01

Agree a family code word

Anyone claiming an emergency must say it. Pick something that is not online, and do the same for payment approvals at work.

02

Hang up and call back

On the number you already have for that person or company. Never the number that called, never one the caller gives you.

03

Shorten your voicemail greeting

Use the carrier's default or a few words. A long personal greeting is a free, clean sample.

04

Drop voice as a password

If your bank offers voice ID, ask for a different second factor. A voice that can be cloned is not a secret.

05

Slow the money

Gift cards, crypto, wires and payment apps to strangers are the scammer's exits. A real emergency survives a ten-minute check.

06

Check the recording

If a voice note or voicemail is driving a decision, save it and run it through the free AI voice detector. Then make the callback anyway.

More defences, including help desk controls and what to do after a payment, are in how to protect yourself from vishing.

Law

Is it illegal to clone someone's voice?

Cloning a voice is not illegal in itself in most places. What you do with it can be. In the United States, using a clone to deceive someone out of money is fraud under existing law. Since February 2024 the FCC has treated AI-generated voices in robocalls as "artificial" under the Telephone Consumer Protection Act, so they need prior consent (FCC). Several states have added rules on voice likeness and on deceptive AI in election material; Tennessee's ELVIS Act of 2024 protects a person's voice as part of their likeness.

Using someone's voice commercially without permission can also breach right-of-publicity law. The rules differ by state and country and are changing quickly. This is general information, not legal advice.

FAQ · Voice cloning

Common questions about voice cloning

What is voice cloning in simple terms?
Voice cloning is AI that copies a specific person's voice from a recording, so software can make that voice say new words. It is used for audiobooks, dubbing and restoring speech, and it is also used by scammers to impersonate relatives, executives and officials on the phone.
How much audio does it take to clone a voice?
Less than most people expect. Research models have produced speech from a few seconds of a speaker, and consumer services clone from a short upload. More clean audio gives a more convincing clone, but a clip from a social video can be enough to fool someone over a phone line.
What is a voice cloning scam?
A voice cloning scam is a call, voice note or voicemail that uses a cloned voice of someone you trust to ask for money, a passcode or access. The most common versions are the family emergency or grandparent scam, executive payment fraud and calls impersonating banks or officials. Hang up and call the person back on a number you already hold.
Can you tell if a voice is cloned just by listening?
Not reliably. Flat emotion, odd breathing, a background that never changes and mispronounced names are clues, but phone compression hides many of them. The FBI says AI-generated content is often difficult to identify. Verify through a callback, and analyse a saved recording with an AI voice detector.
Can an AI voice detector detect a cloned voice?
Yes, from a recording. A detector scores the statistical traces synthetic speech leaves behind. AI Voice Detector returns a four-step verdict, an AI probability score and a per-segment breakdown. It does not screen live calls and does not identify whose voice was cloned, and a human result is not proof that a request is genuine.
Is it illegal to clone someone's voice?
Cloning itself is not illegal in most places, but using a clone to deceive or defraud someone is. In the US the FCC treats AI-generated voices in robocalls as artificial under the TCPA, and some states, such as Tennessee with its 2024 ELVIS Act, protect a person's voice as part of their likeness. This is general information, not legal advice.
How do I stop my voice from being cloned?
You cannot fully prevent it if your voice is public, but you can make a clone useless: agree a family code word, always call back on a known number, keep your voicemail greeting short, avoid voice ID as a bank password, and treat any urgent request for money or codes as suspect until verified.

Think a voice was cloned? Check the recording.

Upload a voicemail, voice note or call recording. Free for a single check, no account needed. Then make the callback anyway.

Check a recording