voice conversion (VC)

in English: voice conversion / speech-to-speech

advancedspreading

An AI method that makes a recorded or live speech sound like another person's voice while preserving the text, rhythm, and intonation.

In plain terms

Voice conversion is like a voice costume: you speak the sentence in your own tempo and intonation, but the listener hears someone else's timbre. The machine does not invent what you say; it only swaps the voice's 'clothing'.

This differs from text-to-speech programs. Those convert written text into speech and decide the intonation themselves. Here, a human performance is the starting point, so laughter, sighs, hesitation, and irony are preserved.

It is used for dubbing and audiobooks, for voicing characters in games, and for patients whose speech has changed or become difficult to understand due to illness or surgery. However, the same technique can be used to impersonate someone else's voice, making it one of the most sensitive generative areas.

How it works

The system breaks down the input speech: it handles separately what you say (content), how you say it (pitch, rhythm, volume), and who says it (timbre).
From a short sample of the target voice, it creates a "voiceprint." This is a number sequence describing the speaker's unique timbre.
A neural network reassembles speech: it overlays the target voiceprint onto your content and delivery style.
A sound-generating module (vocoder) converts this into an audible sound wave. Fast systems do this in a fraction of a second, enabling them to work in live conversations.
Some providers embed inaudible watermarks in the output and tie usage to the voice owner's consent. However, this is not standard practice.

An everyday example

You watch a foreign series in Hungarian and notice that the Hungarian text is spoken in the original actor's voice. A Hungarian voice actor performed the role, and the voice timbre was later replaced with that of the original actor.

Who is affected

everyoneA familiar voice on the phone no longer proves that the person you think is speaking is actually the one talking.

bank customer, investorScammers may request urgent transfers in the voice of a relative or bank officer, and voice-based authentication becomes vulnerable. → Money

patient, doctorFor those whose speech is distorted due to illness or surgery (for example, laryngectomy), the system can transform their hard-to-understand or mechanical voice into a more natural one, even into their own past voice based on earlier recordings. For those who can no longer speak at all, a text-to-speech tool helps. → Health

marketerAds, tutorial videos and multilingual content can be produced faster and more cheaply in a consistent brand voice, but only with the permission of the voice's owner. → Marketing

A common misconception

MythVoice conversion is the same as when a computer reads typed text in someone's voice.

In realityHere, the starting point is not text but a person's actual speech, preserving its intonation and emotion. This makes it sound more natural and harder to detect.

Limits and risks

Mimicking another's voice without their consent may be illegal: the human voice is personal data and is also protected by personality rights.
A recording or phone call alone is no longer reliable evidence; in case of a suspicious request, call the person back on a known number or ask something only they would know.
Quality degrades with noisy recordings, singing, strong emotions, and less common languages. The intonation comes from the original speaker, but models trained mainly on English speech can distort Hungarian sounds, making the result sound foreign.
Watermarks and detection programs are not infallible, and open-source tools often have no protection at all.

Similar, but not the same

text-to-speech – The text-to-speech reader generates speech from written text and chooses the emphasis itself, while voice conversion replaces the timbre of an existing human performance.
voice cloning – Voice cloning is the replication of a specific person's voice from any input; voice conversion is the process that creates speech from speech and often uses a cloned voice as the target.
traditional voice distortion – A voice changer only deepens, raises, or adds effects to the voice, while voice conversion generates a different, believable human voice.
conversational voice model and speech translation (also "speech-to-speech") – The term 'speech-to-speech' is also used for systems that respond to speech with their own reply or translate it into another language. Voice conversion does not reply or translate; it returns the same sentence in a different voice.

Related concepts

Broader

generative AI, speech synthesis, synthetic media

Narrower

real-time voice conversion, singing voice conversion, accent conversion

On the edge

voice cloning, text-to-speech, deepfake, AI watermark, voice-based agent

A brief history

Voice conversion has been researched since the late 1980s, initially with statistical methods. The Voice Conversion Challenge international research competition launched in 2016, and since then neural networks have made the results increasingly lifelike.

Where things stand

As of late 2026, voice conversion is available in commercial services and free open tools, works in live conversations, and is regularly used by the media and gaming industries. Because of voice-based fraud, authorities and banks are also dealing with it. The transparency rules of the EU AI Act (Article 50) apply from August 2026: artificially generated audio must be marked in a machine-detectable way, and for misleading fakes (deepfakes) it must be disclosed that they are artificial. Transitional periods may apply to systems placed on the market earlier.

Status: 2026 November

Also known as

voice conversion · voice conversion · voice conversion · speech-to-speech · STS · S2S · voice changer · AI voice changer

Sources

Wikipedia: Audio deepfake

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.