AI Potluck
Back to Gap Map Model components / Speech & audio

Chatterbox

Resemble AI
open weights / Overall score: 2.8

Resemble AI's open text-to-speech family: zero-shot voice cloning with an emotion-exaggeration control, the original English model, Turbo and Nano variants, and Chatterbox Multilingual V3 across 23 languages. Every output carries Resemble's Perth watermark.

Code and weights are MIT. Resemble's hosted Chatterbox HD is a separate, closed product.

Openness

3 high confidence
3.0
weights
open(checkpoints on the Hub, ungated)
data
described(0.5M hours of cleaned data, not released)
code
partial(inference code only)
license
MIT(OSI)

Chatterbox is MIT-licensed, code and weights alike, which is OSI-approved. Resemble sizes its training data at half a million hours but publishes neither the data nor a training pipeline.

Adoption

4 high confidence
4.0

Hugging Face downloads of the main repository. The chatterbox-tts package is marked as not a separate channel because it loads the Hub checkpoints.

Capability

2 low confidence
2.0

Listeners prefer most commercial voices, and Kokoro, over the original Chatterbox. Its draw is voice cloning and emotion control, which the arena does not score, and its newer models are unrated.

Verified 2026-09-27