Chatterbox TTS

Resemble AI publishes this family of open-weight speech synthesis models along with PyTorch inference code, installable from PyPI as chatterbox-tts. Each model can copy a speaker from a brief audio sample with no fine-tuning step. The lineup covers the original 0.5B English model, which offers exaggeration and classifier-free guidance controls for steering emotion and pacing; a 0.5B multilingual model, now at V3, handling 23 languages; a 350M Turbo model for English voice agents whose distilled decoder runs one step instead of ten and which reads inline tags like [laugh] or [cough]; and a 110M Nano variant aimed at CPUs and on-device use. Six single-language fine-tunes add regional variants such as Latin American Spanish, Brazilian Portuguese, and Hindi. Code runs on CUDA, Apple MPS, or plain CPU and was developed against Python 3.11. Both the repository and the Hugging Face weights use the MIT license, so commercial use is permitted, though every generated clip carries an inaudible Perth watermark that a bundled detector can find later. A strong option for self-hosted narration, dubbing, or voice-agent pipelines.