Chatterbox Multilingual: One Cloned Voice, 23 Languages, No Accent Drift

Clone a voice in English with most multilingual TTS systems, then switch it to Spanish, and something breaks — the accent drifts, the rhythm changes, and the output starts sounding like a different person entirely. Chatterbox Multilingual was built to solve exactly that failure mode: a single open-source model that clones a voice once and holds its identity — timbre, accent, cadence — across 23 languages, without retraining or swapping models per market.

What Is Chatterbox Multilingual?

Chatterbox Multilingual is Resemble AI's general-purpose multilingual text-to-speech model, part of the same open-source, MIT-licensed Chatterbox family as the original English model and the low-latency Turbo variant. The current version, V3, keeps the same 0.5B-parameter architecture as its predecessor while specifically improving three things: speaker similarity across language switches, reduced hallucination (less unwanted repetition or off-prompt continuation), and more natural, conversational delivery in every supported language.

It supports 23 languages out of the box: Arabic, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, and Turkish — covering both widely spoken languages and less commonly served ones like Swahili and Malay in the same release.

The Feature That Actually Matters: Cross-Language Voice Transfer

Most "multilingual" TTS marketing means the vendor ships a different model — or a different fine-tune — per language, and voice identity is never guaranteed to survive the handoff. Chatterbox Multilingual works differently: you provide one reference clip in any language, and the model can generate speech in any of the other 22 while preserving the original speaker's vocal characteristics. Clone an English speaker and have them speak French. Take a Hindi reference clip and generate German. The reference audio's language doesn't need to match the language you're generating — that's the whole point of a single cross-lingual model instead of 23 separate ones.

For projects that need tighter control over a specific market, Resemble AI also offers a Single Language Pack — dedicated fine-tunes for six priority languages and regional variants (Mandarin Chinese, Latin American Spanish, Spain Spanish, Brazilian Portuguese, Portugal Portuguese, and Hindi) — for cases where dialect accuracy or regional pronunciation matters more than broad coverage.

Getting Started with Chatterbox Multilingual

1. Install the package. pip install chatterbox-tts pulls in the full Chatterbox family, including the multilingual class — there's no separate package to track down.
2. Load the multilingual model explicitly. Import ChatterboxMultilingualTTS from chatterbox.mtl_tts (rather than the standard ChatterboxTTS class used for English-only generation) and load it with t3_model="v3" to use the current release.
3. Provide a reference clip and a language ID. Pass a 5–10 second (or longer) audio sample as the voice reference, along with a two-letter language code (fr, ja, ru, and so on) for the target output language.
4. Tune cfg_weight for cross-language generation. Setting cfg_weight to 0.0 minimizes accent bleed from the reference clip's original language — the default 0.5 is tuned for same-language generation, not cross-lingual transfer.
5. Test in the browser before committing to local setup. Resemble AI hosts a Hugging Face Space for the multilingual model, useful for validating voice and language quality before wiring it into a pipeline.

Tips for Better Cross-Language Results

• Match the reference clip's language tag when possible. If cross-language transfer isn't your goal, using a reference clip in the same language you're generating gives more predictable output than relying on cfg_weight adjustments to bridge the gap.
• Use a clean 5–20 second clip. Background noise in the reference audio propagates into every language you generate from it, so it's worth getting the source clip right once rather than troubleshooting per-language artifacts later.
• Reach for the Single Language Pack when a market needs precision. If a specific language is central to your product — not just supported — a dedicated fine-tune will typically outperform the general-purpose multilingual model on dialect accuracy and regional pronunciation.
• Expect some variance early on. Community testing has noted occasional artifacts or accent-transfer inconsistencies depending on the reference clip — treat your first few generations per language as a calibration step, not a final result.
• Budget modest VRAM. Real-world testing has reported multilingual generation running comfortably in the 4GB VRAM range on a modern GPU — solidly accessible hardware for a 23-language model.

Chatterbox Multilingual vs. the Rest of the Family

If your product needs to speak more than one language from a single cloned voice, Multilingual is the model built for that job specifically — the original and Turbo variants are English-only by design.

Frequently Asked Questions

How many languages does Chatterbox Multilingual actually support?

23 languages in the general-purpose V3 model, spanning widely spoken languages like Spanish, Chinese, and Arabic alongside less commonly served ones like Swahili, Malay, and Danish.

Can I clone a voice in one language and generate speech in a different one?

Yes — this cross-language voice transfer is Chatterbox Multilingual's core feature. Provide a reference clip in any supported language and generate output in any other, with the speaker's vocal identity preserved. Setting cfg_weight to 0.0 gives the cleanest cross-language results.

How much reference audio does voice cloning require?

As little as 5–10 seconds, though longer clips (up to around 20 seconds) tend to produce more stable results, particularly for cross-language transfer.

What's the difference between the general multilingual model and the Single Language Pack?

The general model (V3) covers all 23 languages from one set of weights. The Single Language Pack provides dedicated fine-tunes for six priority languages — Mandarin Chinese, three Spanish/Portuguese regional variants, and Hindi — for use cases that need tighter dialect accuracy than the general model provides.

Is Chatterbox Multilingual free for commercial use?

Yes. Like the rest of the Chatterbox family, it's released under the MIT license, and every generated clip includes Resemble AI's built-in PerTh watermarking for provenance.

Try a Comparable Multilingual Voice Cloning Feature

Chatterbox Multilingual itself is available through Resemble AI's open-source implementation and official demo. If you want to explore browser-based multilingual voice cloning, try Voice Clone as a comparable feature. It is not Chatterbox Multilingual and does not run the Chatterbox model, but it lets you test a similar voice-cloning workflow without setting up the model locally.

Only clone voices you own or have explicit permission to use.

Explore Voice Clone →