# Resemble AI releases Chatterbox-Turbo, a 350M MIT TTS built for voice agents

> Chatterbox-Turbo is a 350M English TTS with paralinguistic tags, one-step decoding and voice cloning; MIT licensed and watermarked.

Published 2025-12-02

Canonical: https://shipped.anth.us/open-weights/resemble-ai-chatterbox-turbo/

Resemble AI published Chatterbox-Turbo, a 350M-parameter English text-to-speech model aimed mainly at low-latency voice agents, though the card says it also suits narration. The speech-token-to-mel decoder was distilled from 10 generation steps to one, reducing compute and VRAM compared with earlier Chatterbox models. It supports paralinguistic tags such as [cough], [laugh] and [chuckle], and clones a voice from a reference clip; the card's example uses a 10-second clip. Every output carries Resemble's PerTh watermark, which the card describes as imperceptible and surviving MP3 compression.

The card gives no numeric benchmarks; a Podonos evaluation appears only as an image. Its "sub 200ms" latency figure refers to Resemble's paid service, not this model, and the tuning tips are written for the original Chatterbox. The multilingual Chatterbox (23-plus languages) is a separate model.

The licence is MIT. The card gives no release date; the date used is the Hugging Face repo creation date, 2 December 2025, and the collection was updated 15 December.

MIT, watermarked and English only. The latency number belongs to the paid product.

## Sources
- [Resemble AI model card on Hugging Face](https://huggingface.co/ResembleAI/chatterbox-turbo)
- [Hugging Face repo metadata (creation date)](https://huggingface.co/api/models/ResembleAI/chatterbox-turbo)

## Editor's first pass (AI-assisted)
- Real 4.5: MIT weights on Hugging Face with a clear usage example.
- Useful 3.5: Tags like [laugh] and cloning suit voice agents, English only.
- Notable 3.5: Part of the widely used Chatterbox family.
- Novel 3: Distilling the decoder from 10 steps to one is the main change.