Chatterbox is a family of open text-to-speech models from Resemble AI for zero-shot voice cloning: you pass text and a short reference clip, and the model generates speech in that voice. The family spans a 500M-parameter English model with classifier-free-guidance and exaggeration controls, a 500M Multilingual V3 model covering 23 languages, a 350M Turbo model for low-latency English voice agents, a 110M Nano model for CPU and on-device inference, and a Single Language Pack of dedicated finetunes.
Turbo and Nano use a distilled speech-token-to-mel decoder that generates in one step instead of ten, and both accept paralinguistic tags such as [laugh], [chuckle], and [cough] inline in the text. Nano runs at three times realtime on 8 CPU cores. Multilingual V3 keeps the V2 language coverage while improving speaker similarity across languages and reducing repetition and off-prompt speech. Every generated file carries Resemble's Perth neural watermark, which survives MP3 compression and editing and can be checked with the perth package. Turbo was evaluated on the Podonos platform against ElevenLabs Turbo v2.5, Cartesia Sonic 3, and VibeVoice 7B, with reports published there.
| Model | Size | Languages | Notes |
|---|---|---|---|
| Chatterbox-Turbo | 350M | English | paralinguistic tags, lower compute and VRAM |
| Chatterbox-Nano | 110M | English | Turbo architecture, runs on CPU |
| Chatterbox-Multilingual V3 | 500M | 23+ | cross-language voice cloning |
| Single Language Pack | 500M each | 6 finetunes | zh-cmn, es-mx-latam, pt-br, es-es, pt-pt, hi |
| Chatterbox | 500M | English | CFG and exaggeration tuning |
Features
- Zero-shot voice cloning: a reference clip of about 10 seconds sets the output voice
- Exaggeration and CFG controls: exaggeration and cfg_weight parameters tune expressiveness and pacing on the original model
- Paralinguistic tags: inline [laugh], [chuckle], [cough], and similar markers in Turbo and Nano
- Single-step decoder: Turbo and Nano generate mel output in one step
- Multilingual synthesis: language_id selects among Arabic, German, Spanish, French, Hindi, Japanese, Korean, Chinese, and 15 more
- Single Language Pack: dedicated finetunes for Chinese, Latin American and Spain Spanish, Brazilian and Portugal Portuguese, and Hindi
- CPU inference: Nano targets on-device deployment with tight memory and latency budgets
- Perth watermarking: imperceptible neural watermark in every output with an extraction script
- Python API: ChatterboxTTS, ChatterboxMultilingualTTS, and ChatterboxTurboTTS classes with from_pretrained and generate
