Glossary
What is AI voice cloning
In short
AI voice cloning is the creation of synthetic speech that sounds like the voice of a specific person; the technology is useful for voice-over, but scammers use it too, so you need to know how to protect yourself.
Images, video, audio · voice cloning
MMOOZZGG.com editorsChecked: 2 min read
For a child
Some parrots repeat words so well that it seems grandma is talking. A computer has learned to copy a person’s voice even more exactly. So if someone calls you in a familiar voice and asks you to tell a secret, a password or send money quickly, do not rush. Hang up and ask your mom or dad. You can make up a secret word with your family in advance to check.
In plain words
A neural network listens to recordings of a person’s speech and learns the features of their voice: timbre, pitch, the way they pronounce sounds. Then any text can be voiced as if this person read it. Lawful uses are clear: an actor gives consent and voices an ad in several languages, and a person who is losing their voice to illness saves it in advance. But scammers use the same trick: they call on behalf of a relative or a boss and ask for an urgent money transfer. The rule of protection is simple: do not trust a voice alone, hang up and call back on a number you know yourself.
For business
For a company, this topic has two sides. As a tool, a voice actor’s clone saves studio time for courses, videos and phone menus, but you need a written contract with the person: for which texts, periods and channels they allow their voice to be used. As a threat, it hits payments and access: a scammer calls the accountant in the director’s voice or calls support on behalf of a customer. Protection rests on procedures, not on hearing: urgent payments are confirmed through a second channel, the voice is never the only means of identification, staff get code phrases and take part in training drills. Teams review such scenarios when they assess AI risks before launching voice services.
For a specialist
Voice cloning is speech synthesis adapted to a target speaker. The options: fine-tuning a synthesis model on the speaker’s recordings, transfer from a short sample without fine-tuning (zero-shot), and voice conversion, which changes the timbre of speech already spoken. In VALL-E (Wang et al., 2023), synthesis is framed as language modeling over discrete codes of a neural audio codec; the model was trained on 60 thousand hours of English speech, and the authors report personalized speech from a recording of about three seconds. Protection runs along several lines. Spoofing detection (anti-spoofing): the ASVspoof 2019 database combined synthesis, voice conversion and replayed recordings, and listening tests showed that people cannot tell some fake phrases from real ones. The US regulator FTC ran a challenge for solutions against harmful cloning, and its advice to consumers (March 2023) comes down to checking a call through a known number.
Example
The accountant gets a call from the “CEO”, who asks her to pay a new supplier’s invoice urgently. Following the policy, she hangs up, calls the director back on the known number and learns that he asked no one for anything.
Related terms
Questions and answers
Sources
- Wang et al., Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers (VALL-E), 2023arxiv.org
- Wang, Yamagishi, Todisco et al., ASVspoof 2019: A large-scale public database of synthesized, converted and replayed speech, 2019arxiv.org
- FTC, Voice Cloning Challenge, checked 16 October 2026ftc.gov
- FTC, Scammers use AI to enhance their family emergency schemes, 20.03.2023consumer.ftc.gov
Not sure whether your company needs this?
AI solutions map