Chatterbox TTS: Open-Source Voice Cloning, Explained
Chatterbox TTS is an MIT-licensed open-source voice cloning model from Resemble AI. See the Turbo, Nano and Multilingual variants, watermarking and limits.
Chatterbox TTS is an open-source text to speech model from Resemble AI, released under the MIT license. It clones a voice from a reference clip, comes in variants from 110M to 500M parameters, and embeds a PerTh watermark in every output. Resemble is no longer taking new customers for its hosted service, but it says the models stay open and free to use.
What is Chatterbox TTS?
Chatterbox is a family of open speech models. The GitHub repository is licensed MIT and showed 27.0k stars and 3.6k forks, and the Hugging Face model page showed about 1.67 million downloads last month. Both are snapshots from 2026-10-17. Resemble's own description calls the models "open and free to use".
Launch dates are easy to mix up, so here is the sequence as the sources support it. The original model appears in the Artificial Analysis listing as May 2025. A multilingual version followed, and Turbo was published in December 2025, with its Hugging Face collection updated on 2025-12-15.
Chatterbox models compared
| Variant | Parameters | Languages | Notes |
|---|---|---|---|
| Chatterbox-Turbo | 350M | English | Paralinguistic tags such as [laugh], lower compute, needs a reference clip for cloning |
| Chatterbox-Nano | 110M | English | Runs on CPU. The vendor says 3x faster than realtime on 8 CPU cores |
| Chatterbox-Multilingual V3 | 500M | 23 listed | Vendor says it improves speaker similarity and reduces hallucinations |
| Single-language packs | 500M each | Six packs | Chinese, Latin American Spanish, Brazilian Portuguese, Spain Spanish, Portugal Portuguese, Hindi |
| Original Chatterbox | 500M | English | CFG and exaggeration tuning |
The GitHub list of 23 multilingual languages is: Arabic, Danish, German, Greek, English, Spanish, Finnish, French, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Dutch, Norwegian, Polish, Portuguese, Russian, Swedish, Swahili, Turkish and Chinese. Resemble's marketing page gives a different count, so use the repository list.
How voice cloning works
Cloning starts from a reference clip. The Python API accepts an audio_prompt_path for the clip. The README sets no minimum length and uses a 10-second example file. Resemble's product FAQ says 5 to 20 seconds, so treat that as the vendor's guidance.
The model card gives tuning advice for the two main controls, exaggeration and CFG (classifier-free guidance):
| Situation | Suggested setting | |---|---| | Most prompts | Exaggeration 0.5 and CFG 0.5 (the defaults) | | Fast-speaking reference voice | Lower CFG to about 0.3 | | Expressive speech | CFG about 0.3 and exaggeration 0.7 or higher | | Reference clip in a different language | Match the clip's language to the tag, or set CFG to 0 to limit accent carryover |
Higher exaggeration speeds speech up, so lowering CFG slows it back down. These are the card's recommendations, not guarantees for your voice.
The PerTh watermark
Every output includes Resemble's PerTh watermark. The project describes it as "imperceptible neural watermarks that survive MP3 compression, audio editing, and common manipulations", and it provides a script to check for the mark. The card claims "nearly 100% detection accuracy". That is a vendor claim. The product FAQ says removing the watermark goes against intended use.
The MIT license lets you modify the code, so the watermark is only as strong as your willingness to keep it intact. The contrast is useful. Kyutai chose not to watermark its open model and instead restricts cloning to pre-computed voice embeddings, arguing that watermarks can be disabled in open-source models.
Watermarking is also becoming a compliance topic. The European Commission's code of practice on AI-generated content notes that Article 50 transparency obligations have applied since 2 August 2026, and the code was finalized after a plenary on 10 June 2026. That does not mean Chatterbox satisfies the Act. This is general information, not legal advice.
Install and run
- Install:
pip install chatterbox-tts, or clone the repository and runpip install -e .. - Tested setup: Python 3.11 on Debian 11, according to the README.
- Devices: the
devicesetting acceptscuda,cpuormps. - Cloning: pass the reference clip through
audio_prompt_path.
Nothing here requires an account, but you do need to host it yourself. For what that involves, see self hosted text to speech vs hosted APIs.
Chatterbox vs ElevenLabs: what the evidence shows
Resemble has published a blind test in which Chatterbox beat ElevenLabs. The independent leaderboard does not show the same picture. Both belong in the comparison.
| Source | What it shows | Notes | |---|---|---| | Podonos blind test (commissioned by Resemble) | 63.75% preferred Chatterbox, 16.25% ElevenLabs, 11.25% strongly ElevenLabs, 8.75% no preference | Vendor-reported. Sample size and model version are not shown | | Artificial Analysis Speech Arena, read 2026-10-17 | Chatterbox at Elo 1025 (rank range 71-78). Chatterbox HD, not marked open weights, at 1104 | Independent listener votes. Eleven v4 Turbo at 1330 is the top row |
Read the Podonos figures as a claim Resemble made, and the leaderboard as a dated snapshot. Press coverage of Turbo repeats vendor claims about latency and about beating ElevenLabs and Cartesia. This article does not quote those latency figures, because they describe the vendor's systems and hardware rather than the open model.
What happened to Resemble's hosted service
Resemble published a post dated 2026-08-13 that explains the change. It says "We're supporting our existing customers but not taking new ones." It also says "The voice models we built are open and free to use, but we won't be producing new ones for commercial sale." Chatterbox and related models now sit under "Voice AI Research", and the company's documentation says they are "no longer offered as commercial paid products".
The license FAQ says MIT permits "commercial use, self-hosting, modifying the weights, and production deployment without royalties or usage caps". Resemble has also moved toward deepfake detection, which is the subject of its August post. The practical result is that Chatterbox is an open model you run yourself, not a hosted service you can buy.
Cloning rules: consent and laws
The Chatterbox model card includes a short ethics line: "Don't use this model to do bad things." Clone only voices you have permission to use. Laws on cloning vary by country and by use, so check the rules that apply to you. This section is general information, not legal advice.
Chatterbox alternatives
- Qwen3-TTS: the base models offer a "3-second rapid voice clone" and are Apache-2.0 licensed, released 2026-01-22.
- XTTS-v2: clones from a 6-second clip under the Coqui Public Model License. The model page does not state commercial terms.
- Hosted custom voices: a studio service removes the setup work. OndaLabs custom voices let you clone from a recording you have permission to use, or design a voice from a text description. Voice design or cloning costs 1,500 credits. Custom voice slots depend on the plan: Creator $22 includes 3, Pro $49 includes 10, and Business $199 includes 30. Pricing details are on the pricing page, and text to speech covers the rest of the studio.
Frequently asked questions
What is Chatterbox TTS?
Chatterbox TTS is an open-source text to speech and voice cloning model from Resemble AI. It is released under the MIT license and comes in several variants, including Turbo, Nano and Multilingual. Every output carries the PerTh watermark.
Is Chatterbox free for commercial use?
The MIT license permits commercial use, self-hosting and production deployment without royalties or usage caps, according to Resemble's license FAQ. Your own use still needs to follow the consent and legal rules that apply to you. Check the license file in the repository before you ship.
How much audio does Chatterbox need to clone a voice?
The README sets no minimum. Resemble's FAQ says 5 to 20 seconds, and its headline says about 5 seconds. Start with a clean clip in the same language as your target text, then adjust exaggeration and CFG.
What languages does Chatterbox support?
The original and Turbo models are English only. The Multilingual V3 model lists 23 languages on GitHub, and single-language packs cover Chinese, Latin American Spanish, Brazilian Portuguese, Spain Spanish, Portugal Portuguese and Hindi.
Does Chatterbox watermark its audio, and can I remove it?
Yes, every output includes the PerTh watermark, which Resemble says survives MP3 compression and editing. The MIT license lets you modify the code, so the mark can be removed by someone who changes the software. Removing it goes against the vendor's intended use.
Is Chatterbox really better than ElevenLabs?
Resemble's blind test favored Chatterbox, but the test was commissioned by Resemble and its sample size is not shown. The independent Artificial Analysis snapshot read on 2026-10-17 puts Chatterbox at 1025 and ElevenLabs' top model at 1330. Neither result settles the question for your voice and text.