Chatterbox is a family of open-source, state-of-the-art text-to-speech (TTS) models from Resemble AI, released under a permissive MIT license and now past 25,000 GitHub stars. Built around zero-shot voice cloning, it can reproduce a target voice from only a few seconds of reference audio and generate natural, expressive speech from text — all with a `pip install chatterbox-tts` and a handful of Python lines. Rather than shipping a single model, the project has grown into a small zoo of variants tuned for different trade-offs between quality, latency, and language coverage, which is a large part of why it has become one of the more widely adopted open TTS stacks. ## Zero-Shot Voice Cloning The core promise of Chatterbox is that you do not need to train a voice. You supply a short reference clip — roughly ten seconds is typical — and the model synthesizes new speech in that voice for arbitrary text. This zero-shot approach removes the dataset-collection and fine-tuning steps that traditionally gated custom TTS, making it practical to spin up a new voice on the fly for prototypes, agents, or narration. The models run on CUDA GPUs but also support CPU and Apple's MPS backend, so they are approachable on a range of hardware. ## A Model Zoo for Different Needs Chatterbox splits into purpose-built variants. Chatterbox-Turbo is a streamlined 350M-parameter English model aimed at low-latency voice agents, with a distilled decoder that cuts generation from ten steps to a single step while keeping high-fidelity output. Chatterbox-Multilingual V3, at 500M parameters, targets broad language coverage — 23 or more languages — with improved speaker similarity and reduced hallucination for cross-language cloning. A Single Language Pack adds dedicated finetunes for priority languages where dialect-specific quality control matters. This lets teams pick the point on the speed/quality/coverage curve that fits their use case instead of forcing one model to do everything. ## Expressive Control Beyond raw synthesis, the models expose creative controls. The Turbo model supports native paralinguistic tags such as `[laugh]`, `[cough]`, and `[chuckle]` inline in the text, letting you inject realistic non-verbal cues. The original Chatterbox exposes classifier-free guidance and an exaggeration parameter for tuning delivery, which makes it useful for narration and creative workflows in addition to conversational agents. ## Trade-offs and Limitations High-quality, low-latency generation still benefits from a GPU, and multilingual cloning quality varies by language and by the quality of the reference clip. As with any capable voice-cloning system, the ability to reproduce a person's voice from seconds of audio raises clear consent and misuse concerns; Resemble AI ships watermarking guidance, but responsible use is left to the deployer. Finally, the fast-moving model zoo means tutorials and API names can drift between releases, so pinning versions is wise for production. ## Who Should Use This Chatterbox is a strong fit for developers building voice agents, dubbing and localization pipelines, accessibility tools, or creative audio projects who want production-grade TTS without training a model or paying per character. The MIT license and active community lower the risk of building on it, while Resemble AI's hosted service offers a scaling path for teams that later need sub-200ms latency at volume.