Chatterbox
Open text-to-speech models with zero-shot voice cloning, now an unmaintained archive.
⬦ Free and open source
Released under the MIT license. This page carries no affiliate link. The button goes to the official project.
Get Chatterbox →Chatterbox is a family of open text-to-speech models from Resemble AI that clone a voice from a short reference clip (the examples use a 10-second one), with no training step. Turbo covers low-latency English, Multilingual V3 covers 23+ languages, and Nano is small enough to run on a CPU.
The caveat: as of August 2026 the repository describes Chatterbox as retired research that is no longer maintained, kept public as an MIT-licensed archive. The models still run, but nothing will be fixed. Every file it generates also carries Resemble AI’s PerTh watermark. To size a GPU for the larger models, see local AI by VRAM; the setup is in the guide to self-hosting voice cloning with Chatterbox.
The good
- MIT licensed
- Zero-shot voice cloning from a reference clip
- Turbo (350M, English) for low latency, Multilingual V3 (500M) for 23+ languages
- Nano (110M) targets CPU and on-device use
The catch
- Retired research since August 2026: no support, updates or paid tier
- Every output carries an imperceptible PerTh watermark
- Setup examples target a CUDA GPU; Nano is the CPU option
Pricing
Free and open source.
Alternatives to consider
- Kokoro-82M A tiny 82M-parameter TTS model with quality above its weight.
- Piper Fast, lightweight neural text-to-speech that runs offline.
- ElevenLabs The realistic-voice benchmark: TTS, voice cloning, and dubbing.