Cloning your voice into another language: why the accent carries over, and when you want it to

A cloned voice keeps its accent in every language

A cloned voice speaks a new language with the accent of the person in the reference recording. Clone an American English speaker and have the voice read Spanish, and you hear an American speaking Spanish. Clone a Hindi speaker and have it read English, and you hear Indian English. The words are right; the accent is the speaker's.

Vendors describe their own systems this way. ElevenLabs' help centre puts it plainly: "The language is determined by the text, the accent and pronunciation is determined by the voice itself" (ElevenLabs). Smallest AI's documentation gives a concrete case, saying that a clone of a South Indian English speaker "will retain that accent when generating Hindi or Tamil output" (Smallest AI).

Pāṇini, Sonex Labs' text-to-speech model, works the same way. Its clone endpoint takes reference audio only, with no language field, and the language comes from the text at synthesis time. The public API exposes the voice, the speed and the output format, and nothing that changes accent; a language field exists in the request, but the API documentation says it has no effect on pronunciation. So on Pāṇini the accent cannot be turned down.

Why the accent travels with the voice

The reason lies in how modern cloning works, and it explains both why the accent comes along and why separating it is hard.

One sentence, three reference speakers

These samples show the mechanism with platform voices, which are themselves made from reference recordings. Each reads the same Spanish line: "Hi, I'm calling to confirm your appointment tomorrow at four. Does that still work for you?" Sofia was recorded by a Latin American Spanish speaker, Liz by an American English speaker and Priya by a Hindi speaker. Listen for the vowels and the r sounds.

The last two samples do the reverse, reading an English line with a Hindi voice and an American voice. There is no "reduced accent" version to compare, because Pāṇini has no setting that produces one. What you hear is what a cloned voice from each kind of speaker would sound like.

When you want the accent to carry over

Often the accent is the point. A creator voicing their own videos in Hindi or Spanish wants viewers to hear the person they already follow, and a native-sounding stranger would break that. Accent is part of who the speaker is, and audiences notice when it disappears. Our Hindi AI dubbing lessons look at how creators weighed this on real projects, and the Sonex page for creators covers narration workflows.

The same goes for a brand or founder voice used across markets, and for anyone using a clone as a personal voice, for example after losing the ability to speak. In these cases the accent is identity, and keeping it is honest: listeners hear the real person, speaking a language they may not speak fluently.

An accent can also match the audience. A receptionist for an Indian clinic in Dubai or London may serve callers better with Indian English than with an American voice. A cloned voice can be used by a voice agent in the same way as a platform voice, and Pāṇini is trained for 8 kHz phone audio. Phone lines remove high frequencies, though, so listen to a clone on a real call before launch, not only on headphones.

When you don't: start from a native speaker

If listeners should hear a native speaker, the reference recording has to come from one. Smallest AI's advice is to "record reference audio in the same language as your intended output" (Smallest AI), and ElevenLabs likewise points users to a voice originally recorded in the target language. In practice that leaves three routes, each with a trade-off:

• Use a native platform voice where one exists. Pāṇini has native voices for languages such as Hindi, Tamil, Telugu, Bengali, Marathi, Japanese, English, French, German, Latin American Spanish and Modern Standard Arabic, listed on the text to speech page. It will not be your voice.
• Clone a native speaker of the target language, with their explicit permission. You get a native accent, but again not your voice, and the speaker should know how it will be used.
• Record your reference in the target language. The model then copies how you speak that language, accent included, so this helps only if your accent in it is the one you want.

Consent and ownership of a cloned voice

Cloning a voice needs the speaker's agreement, and it matters more when the clone will speak words the speaker never said, in a language they may not know. Before Sonex Labs creates a voice, its dashboard asks you to confirm that the recording is your own voice, or that you have the speaker's "explicit permission" to create a synthetic voice from it.

You also confirm that you will not use the voice to impersonate anyone, to mislead or for any unlawful purpose, and that you accept responsibility for everything produced with it. The recording and the voice profile are stored for your workspace and used only to synthesise speech for it, and you can delete the voice at any time. A celebrity's or colleague's voice is no exception: clone it only with their explicit permission.

Law sits on top of these terms and varies by country. Our guide to voice-cloning law in India explains the position there. If you clone a native speaker for another market, put the permission in writing and say which languages and uses it covers.