Coqui XTTS: Clone One Voice Into 16 Languages
One short clip, sixteen languages. How to feed Coqui XTTS a clean reference, write scripts each language can pronounce, and test your cloned voice with ten copy-ready scripts.
By Eftal Yurtseven · Nov 17, 2026 · 9 min read

Translating a script is the easy part. Making it sound like the same person said it in Spanish, Turkish and Korean is where most voice projects fall apart: three languages, three strangers. Coqui XTTS was built for exactly that gap. Hand it one short clip of a real voice, pick a target language, and it speaks your text in that language with the timbre of the clip. The trick is that the results depend far less on the model than on two things you control: the reference you feed it and the way you write each script.
This guide is about those two things. You'll get the language codes XTTS-v2 accepts on each::labs, a recipe for the reference clip, rules for writing scripts that each language can actually pronounce, and ten copy-ready test scripts you can paste straight into the text field.
What Coqui XTTS Does Across Languages
XTTS-v2 is a voice generation model from Coqui. Its official model card lists the headline features plainly: voice cloning from a 6-second audio clip, emotion and style transfer by cloning, cross-language voice cloning, multilingual speech generation, and a 24 kHz sampling rate. The card also says v2 added Hungarian and Korean over the first version, along with better prosody and stability.
Cross-language cloning is the part people underestimate. Your reference clip can be English and your output can be Polish. The model carries the voice across and lets the target language set the pronunciation. That means one recording session can feed a whole localization run, which is a very different job from hiring a voice actor per market.
A quick note on the maker. Coqui, the company, closed in early January 2024 after discontinuing its paid web app and API in December 2023, as Analytics India Magazine reported at the time. The XTTS-v2 weights and model card stayed published under the Coqui Public Model License (CPML), and the model runs on each::labs today. Read the CPML before you put generated audio into commercial work.

The 16 Language Codes You Can Pick on each::labs
The model card lists 17 languages. The each::labs language menu exposes 16 of them, and the one missing is Japanese, so plan around that if Tokyo is on your list. Here's what you can select, by the code the API expects:
English (en), Spanish (es), French (fr), German (de), Italian (it), Portuguese (pt), Polish (pl), Turkish (tr), Russian (ru), Dutch (nl), Czech (cs), Arabic (ar), Chinese (zh), Hungarian (hu), Korean (ko) and Hindi (hi).
Two details save headaches. Chinese is zh on each::labs, not the zh-cn you'll see in Coqui's own code samples. And the language field defaults to en, so if you paste a German script and forget to switch it, the model will try to read German with English rules. The result is a confident, fluent mess. Set the code first, paste the text second.

Feed It One Clean Reference Clip
The speaker input is the only required field, and it takes a URL to an audio file. That file is the whole voice. Everything the model knows about the person comes from it, including the room, the mic, the hum of the laptop fan and the mood of the speaker that day.
Coqui's model card says 6 seconds is enough to clone, and its TTS documentation notes that the model uses 12 seconds of reference audio for conditioning by default. So a clip of 6 to 12 seconds of continuous, clean speech covers what the model actually listens to. Longer is not better if the extra seconds are noise.
Record it like this. One speaker, no music bed, no reverb from a tiled bathroom. Speak at the energy you want the output to have, because the model card's "emotion and style transfer by cloning" cuts both ways: a tired, mumbled reference produces a tired, mumbled French narrator. Trim silence from the start and end. If you have a choice, use a clip where the speaker says full sentences rather than a list of words, since natural phrasing gives the model more of the voice's rhythm.
Does the reference language matter? Less than you'd think, since cross-language cloning is the point. Pick the cleanest clip you have, whatever language it's in. What matters is clarity and character, not which words were spoken.
Set the Four Inputs, Then Leave Them Alone
The each::labs schema for XTTS-v2 has four inputs, and that's refreshing. There's no wall of sliders to fiddle with.
text is the script. speaker is the reference clip URL. language is one of the 16 codes above and defaults to English. cleanup_voice is a toggle, on by default, that the schema describes as refining and improving the quality of the generated speech.
Leave cleanup on for your first pass. Turn it off only as an A/B test, when you want to hear whether it's smoothing away a texture you actually like in the voice (a slight rasp, a breathy edge). Change one thing at a time. If you swap the reference and flip cleanup in the same run, you won't know which one fixed or broke the take.
Because the model page sits behind the same API as every other model in the text-to-voice catalog, a localization run is just a loop: same speaker URL, same cleanup setting, a new text and language code on each call. If you want that loop as a reusable pipeline with a translation step in front, build it once in each::labs flows and stop copying and pasting.

Write the Script for the Target Language, Not the Source
Most bad multilingual takes are bad scripts. The model reads what you give it. If the translation is stiff, the voice will be stiff in a perfect accent.
Use the native script, always. Arabic in Arabic letters, Hindi in Devanagari, Russian in Cyrillic, Chinese in Chinese characters. Romanized text in a non-Latin language confuses the pronunciation rules the language code is supposed to apply.
Punctuate for breath. Coqui's documentation explains that long inputs are split into sentences by default, which lets you send long text but can lose some context between sentences. So write sentences that stand on their own. A comma tells the model to pause; a full stop tells it to land. Chinese and Arabic have their own punctuation marks (。 , ؟), and using them helps the rhythm.
Spell out what's ambiguous. Numbers, dates, currencies, units and abbreviations read differently in every language. "3/4" could be a date, a fraction or a score. Write "three quarters" or "tres cuartos" and remove the guess. Do the same with brand names: if your product name should sound English inside a Turkish sentence, test it, and if it doesn't, write it phonetically for that market.
Keep test lines short. Two or three sentences show you everything about a voice in a new language, and a short take is quick to judge. Save the full-length script for after you trust the setup.
Ten Test Scripts to Copy
Each script below is written natively for its language, short enough to judge in one listen, and built to stress something specific: a question, a number, a list, a change in mood. Run them all with the same reference clip and you'll have a fast map of where your voice travels well. Set the language code shown before each one.
English (en), your baseline. Run this first so you know what "sounds like them" means.
Welcome back. This is a short test of my cloned voice, recorded on a quiet Tuesday morning. If it sounds like me, the next twenty minutes are going to be easy.
Spanish (es), a calm, slower read with an imperative at the end.
Bienvenidos de nuevo. Hoy vamos a hablar de algo sencillo: cómo empezar el día con calma. Respira hondo, abre la ventana y escucha.
French (fr), a retail announcement with a spelled-out time.
Bonjour à tous. Notre boutique ouvre ses portes à neuf heures, et la première commande arrive avec un petit cadeau. Merci de votre patience.
German (de), a direct question in the formal register. Listen for the rising intonation.
Willkommen zurück. Bevor wir anfangen, eine kurze Frage: Haben Sie das Handbuch schon gelesen? Wenn nicht, ist das völlig in Ordnung.
Turkish (tr), a creator intro with a count and a punchline.
Merhaba, hoş geldiniz. Bugün size üç kısa ipucu anlatacağım. İlki çok basit: sabah kahvenizi içmeden telefonunuza bakmayın.
Russian (ru), an evening narrator tone with a soft twist at the end.
Добрый вечер. Сегодня мы поговорим о том, как звучит один и тот же голос на разных языках. Это интереснее, чем кажется.
Arabic (ar), a podcast opener. Check the pacing around the comma.
مرحبًا بكم من جديد. في هذه الحلقة سنتحدث عن السفر بحقيبة صغيرة فقط. الفكرة بسيطة، لكنها تغيّر كل شيء.
Chinese (zh), a product explainer with a question and native punctuation.
大家好,欢迎回来。今天我们只讲一件事:怎样用一分钟介绍你的产品。准备好了吗?我们开始吧。
Hindi (hi), a warm, conversational welcome in Devanagari.
नमस्ते, फिर से स्वागत है। आज हम एक छोटी सी बात करेंगे: सुबह की शुरुआत शांति से कैसे करें।
Korean (ko), a polite tutorial intro. Korean is one of the two languages v2 added, so it's worth hearing early.
안녕하세요, 다시 오신 것을 환영합니다. 오늘은 짧지만 유용한 팁 세 가지를 알려드릴게요.
Got a native speaker on the team? Hand them the take, not the script. They'll catch a wrong stress in two seconds that you'd never hear in ten listens.

Read the Results Like an Editor
Judge every take on two separate questions, in this order. First: is this the same person? Play the English baseline and the new take back to back. Timbre and texture should match even when the accent changes. Second: would a native listener accept it? Rhythm, stress, and how the questions rise matter more here than any single sound.
When the voice drifts, the fix is almost always the reference. A clip with background noise or a strong room tone pulls the output toward that sound in every language you render. Re-record or re-trim before you touch anything else.
When pronunciation drifts, the fix is almost always the script. Check that the language code matches the text, swap numerals for words, and split long sentences. If a single word keeps coming out wrong, rewrite it the way a native speaker would spell it out loud. It feels like cheating. It works.
And when a take is simply flat, look at the energy of the reference clip. Style transfer means the model borrows the delivery as well as the voice, so a livelier reference gives you a livelier take in every language at once.
When to Pick Another Voice Model
Coqui XTTS earns its place when the job is one specific human voice carried into many languages. If you don't need a particular person's voice, other models on each::labs may fit better.
For fast, lightweight narration with built-in voices, Kokoro 82M is a small model that punches well above its size; our Kokoro 82M write-up covers how to get the most from it. For expressive, studio-polished delivery, MiniMax's Speech 2.8 HD is the one to try. For characters and interactive voices, Inworld TTS 1.5 from Inworld is worth a listen. Our roundup of the best AI text-to-speech models in 2026 compares them side by side.
If you're new to XTTS and want the basics of cloning in one language before going multilingual, start with our six-second voice cloning guide and come back here when you're ready to travel.
Frequently Asked Questions
How many languages does Coqui XTTS support?
Seventeen on the official model card, sixteen on each::labs. The each::labs language menu covers English, Spanish, French, German, Italian, Portuguese, Polish, Turkish, Russian, Dutch, Czech, Arabic, Chinese, Hungarian, Korean and Hindi. Japanese, the seventeenth, isn't in the menu.
Does the reference clip need to be in the target language?
No. Cross-language voice cloning is one of the model card's listed features, so an English clip can drive a Hindi or Korean take. Clean audio matters far more than the language of the clip.
How long should the reference audio be for XTTS-v2?
Six seconds is the floor the model card names. Coqui's docs say the model uses 12 seconds of reference audio for conditioning by default, so 6 to 12 seconds of clean, continuous speech is the practical target.
What happened to Coqui, and can I still use Coqui XTTS?
The company closed in early January 2024, but the model didn't disappear with it. XTTS-v2 remains published under the Coqui Public Model License, and you can run it on each::labs through the XTTS-v2 model page. Check the CPML terms before using output commercially.



