A detailed guide to Resemble AI Chatterbox Multilingual V3, covering zero-shot voice cloning, multilingual TTS, emotion control, open-source deployment and PerTh watermarking.
Most high-quality voice cloning tools are closed, subscription-based services. Chatterbox Multilingual takes a different route. It is an open-source text-to-speech model from Resemble AI that clones voices from short samples, speaks 23+ languages, and ships with built-in watermarking for responsible use.
This guide explains what Chatterbox Multilingual is, how its voice cloning and voice conversion features work, how to get started, and what to consider on consent, licensing, and cost. To compare it with other tools, browse our Voice Cloning & Voice Conversion category.
Chatterbox is a family of open-source text-to-speech models developed by Resemble AI. The Multilingual line is its general-purpose model for many languages. The latest release, Chatterbox Multilingual V3, is a 0.5B-parameter model that supports 23+ languages, including Arabic, English, Spanish, French, German, Japanese, Korean, and Chinese.
Resemble AI describes Chatterbox as the first open-source TTS model to combine emotion exaggeration control with multilingual zero-shot voice cloning. The V3 release improves speaker similarity and accent preservation across languages, reduces hallucinations, and sounds more natural. The code and weights are available on the official GitHub repository and on Hugging Face, and the project is listed under the MIT license.
Voice cloning (zero-shot) means you give the model a short reference clip and it speaks new text in that voice. No training run is needed.
Voice conversion means you take existing recorded speech and re-speak it in a target voice, keeping the original timing and delivery. Chatterbox includes a voice conversion script for this.
Together they cover two common workflows: generate narration from text in a cloned voice, or transform an existing recording into another voice.
| Capability | What it does |
|---|---|
| Zero-shot voice cloning | Clones a voice from a short, clear reference clip |
| One model for global content, including cross-language cloning |
| Emotion exaggeration control | Dial expressiveness up or down for more dramatic or calmer delivery |
| Voice conversion script | Converts recorded speech into another voice |
| Single Language Packs | Six dedicated fine-tunes for priority languages and regional variants |
| PerTh watermarking | Imperceptible neural watermarks embedded in generated audio |
| Open weights | Download, run, and integrate the model yourself |
| Chatterbox-Turbo | A separate, efficient model for low-latency English voice agents |
The Single Language Packs currently cover Chinese, Latin American Spanish, Spain Spanish, Brazilian Portuguese, Portugal Portuguese, and Hindi.
Chatterbox suits developers, researchers, and technical creators who want an open, self-hostable voice model with no per-minute fees. It is less convenient if you want a polished point-and-click studio, because the open model is a code-first tool. Hosted offerings from Resemble AI and other platforms are better for non-technical teams.
| Need | Better fit |
|---|---|
| Open-source, self-hosted multilingual cloning | Chatterbox Multilingual |
| Low-latency English voice agents | Chatterbox-Turbo |
| Polished web studio with a large ecosystem | ElevenLabs |
| Fast cloning with a hosted web app and open models | Fish Audio |
1. Confirm you have the rights.
Clone only your own voice or one you have explicit permission to use.
2. Prepare a clear reference clip.
The project suggests a clean sample of roughly 5–30 seconds. Use a quiet room, one speaker, and no music or echo.
3. Install the package.
Clone the repository and install it, or use the package name `chatterbox-tts` with pip. Follow the setup notes in the GitHub README.
4. Load the multilingual model.
Choose Chatterbox Multilingual V3 for broad language coverage. For a priority language, consider the matching Single Language Pack.
5. Generate speech with your reference clip.
Provide your text, the language, and your reference audio. Voice cloning requires a reference clip, and without one the model uses a default voice.
6. Tune expressiveness.
Adjust the emotion exaggeration setting to match your content: lower for calm narration, higher for dramatic delivery.
7. Try voice conversion.
Use the voice conversion script to turn a recorded performance into the cloned voice.
8. Test and review.
Listen to every language you plan to publish, since quality can vary. Check names, numbers, and technical terms.
You can also try the model without installing anything in the official Hugging Face demo Space.
Every file generated by Chatterbox includes Resemble AI's PerTh (Perceptual Threshold) watermark. According to the project, these imperceptible watermarks survive MP3 compression and common edits and can be detected with high accuracy. This helps people verify whether audio was AI-generated.
Watermarking does not replace consent. Follow these rules:
This guide is not legal advice. Consult a qualified professional for commercial or high-risk projects.
The Chatterbox model is open source, so there are no license or per-minute fees. Your real cost is the compute you use to run it: a suitable GPU or a cloud service, plus engineering time for setup and maintenance. If you prefer a managed service, check Resemble AI's official site for current hosted and enterprise options, because pricing and features there change over time.
Chatterbox Multilingual is one of the most capable open options for voice cloning today. It combines zero-shot cloning, 23+ languages, expressive control, voice conversion, and built-in watermarking in a model you can run yourself. For developers and technical creators, that means no usage meter, full control over your data, and room to customize.
The trade-offs are practical. You need to manage hardware and setup, quality varies by language, and there is no polished studio interface out of the box. Responsibility for consent and disclosure also remains with you. Start with a short, clean sample of your own voice, test a few languages, and scale once the results meet your standards. Ready to try it?
It is an open-source text-to-speech model from Resemble AI that supports zero-shot voice cloning in 23+ languages, with emotion control and built-in watermarking. See the official GitHub repository.
The model is open source and listed under the MIT license, so there are no license fees. You pay only for the compute you use to run it, or for any hosted service you choose.
Chatterbox Multilingual V3 supports 23+ languages, including Arabic, English, Spanish, French, German, Japanese, Korean, and Chinese. Try it in the official Hugging Face demo.
The project recommends a clear reference clip of roughly 5 to 30 seconds.
Yes. It includes a voice conversion script that re-speaks recorded audio in a target voice.
PerTh is Resemble AI's neural watermarking system. It embeds imperceptible markers in generated audio so the audio can later be identified as AI-generated.
Yes. The model supports emotion exaggeration control, so you can make delivery calmer or more dramatic.
Only with their explicit permission. Using a cloned voice to impersonate or mislead people is unethical and may be illegal.
