Chatterbox
A family of open-source text-to-speech models from Resemble AI, with zero-shot voice cloning. Part of Open-source motion studio: 20 repos.
Checked 2026-10-27 against the GitHub repository, its README, and the npm or PyPI registry where relevant.
| Runtime | Python (developed on 3.11 and Debian 11); device can be cuda, cpu or mps. Local, no per-character fees. |
| How an agent drives it | Python API and a pip package; fits a scripted narration step (text in, WAV out). Output is not bit-identical between runs unless seeded; not tested here. |
| License | Code: MIT (GitHub license file). Model-weight license: not stated in the README sections read; check the Hugging Face model cards before commercial use. |
| Maintenance | Latest GitHub release v0.1.2, 2025-06-13, but the repo was last pushed 2026-07-21 (Nano update) and PyPI chatterbox-tts 0.1.7 is dated 2026-03-26. Active, with releases made through commits and PyPI rather than GitHub releases. |
What it gives you
- Models listed in the README: Chatterbox-Turbo (350M, English, paralinguistic tags such as
[laugh]and[cough], lower compute and VRAM), Chatterbox-Nano (110M, English, runs on CPU at about 3x realtime on 8 cores per the README) and Chatterbox Multilingual V3 (about 500M, 23+ languages). - Input: text plus an optional reference voice clip. Output: audio. Install with
pip install chatterbox-tts. - Every generated file carries Resemble AI’s Perth neural watermark (README), with a script for extraction.
Key differentiator
A free, local voice-cloning model with paralinguistic tags and a built-in watermark, with a small CPU-capable variant. Source: README. Comparisons against ElevenLabs in the yt-base videos are creators’ opinions.
Pick something else when
ElevenLabs (commercial) for managed quality and support. Kokoro (Apache-2.0, 82M) for a small fixed-voice model. F5-TTS (MIT) for another cloning model. Piper for fast offline voices (the original repo is archived; the maintained fork is GPL-3.0). See Text-to-speech and Voice cloning.
Agent fit
Strong. A Python call that turns a script line into audio, runnable in a pipeline with no GUI.
Overlap within the 20
Pairs with faster-whisper (generate the voice, then transcribe it for captions) and Tone.js (music and effects, not speech).