What it is
Chatterbox is a family of open text-to-speech models from Resemble AI, released under the MIT licence. The current flagship, Chatterbox Multilingual V3, has 0.5 billion parameters on a Llama backbone and covers 23 languages, with separate fine-tunes for Chinese, Hindi and regional Spanish and Portuguese. Chatterbox-Turbo, a 350-million-parameter English model released in December 2025, targets low-latency voice agents and accepts tags such as [laugh] and [cough]; Chatterbox-Nano, at 110 million, runs on a CPU.
All variants clone a voice from a short reference clip, and the original model adds an 'exaggeration' control for emotional intensity. Resemble says the models were trained on 0.5 million hours of cleaned audio. The main Hugging Face page records about 1.8 million downloads in the past month.
Our review
Chatterbox is picked for its licence and its watermark, not its sound. On Artificial Analysis's arena the open model ranks 78th of 98 at 1,024 Elo, below Kokoro's 1,065, while Fish Audio's Drama 3 Preview leads open models at 1,248. Resemble's hosted Chatterbox HD fares better at 1,109. The model card's claim that Chatterbox beats ElevenLabs rests on Resemble's own Podonos comparison, not an independent arena.
On misuse it does more than most open models. Every clip carries Resemble's Perth watermark, which the company says survives MP3 compression and editing with nearly 100% detection, and the repository ships a short script that reads it back as 0.0 or 1.0. The protection has limits: the watermark is applied in open code anyone can edit, the only usage policy is a one-line disclaimer ('Don't use this model to do bad things'), and nothing checks consent before a voice is cloned.
The family now covers a wider range of hardware than most open TTS projects. Nano, at 110 million parameters, runs three times faster than real time on eight CPU cores; Turbo, at 350 million, distils the decoder from ten steps to one for voice agents; Multilingual V3 handles 23 languages at 500 million. One trap for engineers: the multilingual loader still defaults to the older V2 checkpoint unless you pass t3_model="v3".
The latency story is thinner than the marketing suggests. The 'sub 200ms' figure on the model cards describes Resemble's paid hosted service, and no latency or VRAM figure is published for the open weights. Cloning across languages also needs care: if the reference clip's language differs from the target, the output picks up the clip's accent unless CFG is set to zero.
Where it earns its place
A voice-agent startup running English phone bots on its own GPUs
Speak replies in a cloned brand voice without per-character fees
Run Turbo, whose one-step decoder is built for agents, from a reference clip of about ten seconds. Self-hosting has no per-character charge; at Artificial Analysis's $40 per million characters for hosted Chatterbox HD, 50M characters a month would cost $2,000.
A trust-and-safety lead at a media company using synthetic narration
Prove which audio clips came from the company's own pipeline
Run the Perth detector over outbound and disputed audio; it returns 1.0 for watermarked files. Because the watermark is applied in open code, it marks honest use rather than stopping a determined forger.
An e-learning company localising courses into Hindi, Spanish and Portuguese
Keep one narrator's voice across languages
Load Multilingual V3 with t3_model="v3", or the dialect fine-tunes for Hindi, Latin American Spanish and Brazilian Portuguese, and set CFG to zero when the reference clip is in another language to avoid accent bleed.
A developer building an offline reading app for laptops
Generate speech on CPU without a server
Chatterbox-Nano loads through the Turbo class with nano=True and runs three times faster than real time on eight CPU cores, per Resemble. It is English-only, so multilingual apps still need the 500M model.
Before you commit
- Setup
- pip install chatterbox-tts; Resemble tests on Python 3.11 and Debian 11, and the models run on CUDA, CPU or Apple MPS. Cloning needs a reference clip; the example uses about ten seconds.
- Integrations
- Python package and Hugging Face weights; fal is listed as an inference provider, and Resemble sells a hosted Chatterbox HD with lower latency.
- Data and privacy
- Self-hosted, so text and reference audio stay local. MIT licence with no revenue limits; the only usage policy is 'Don't use this model to do bad things', and there is no consent check on cloned voices.
- Cost at scale
- Free weights; the cost is compute, and Nano runs on CPU. Artificial Analysis lists hosted Chatterbox at $25 and Chatterbox HD at $40 per million characters, against $0.65 for hosted Kokoro.
- Limits
- No published latency, VRAM or maximum text length for the open models. Turbo and Nano are English-only, and mismatched reference and target languages cause accent bleed unless CFG is zero.
What's good
- MIT licence, commercial use included
- Perth watermark on every clip, with a detection script
- Nano (110M) runs 3x faster than real time on 8 CPU cores
- 23 languages in Multilingual V3, plus dialect fine-tunes
What's not
- 78th of 98 on Artificial Analysis's arena for the open model
- No consent check before cloning a voice
- No published latency or VRAM figures for the open models
- Multilingual loader defaults to V2 unless told otherwise
Best for
Teams that need commercially usable open voice cloning with provenance marking and can accept mid-table voice quality.
Alternatives
Chatterbox: common questions
- Is Chatterbox worth it?
- Chatterbox is the open voice cloner a compliance team can sign off, thanks to its watermark, though not the one a listener would choose. We rate it 3.7 out of 5.
- How much does Chatterbox cost?
- Free under the MIT licence. Artificial Analysis lists hosted Chatterbox at $25 and Resemble's Chatterbox HD at $40 per million characters.
- What are the downsides of Chatterbox?
- 78th of 98 on Artificial Analysis's arena for the open model; no consent check before cloning a voice; no published latency or VRAM figures for the open models; multilingual loader defaults to V2 unless told otherwise.
- Who is Chatterbox best for?
- Teams that need commercially usable open voice cloning with provenance marking and can accept mid-table voice quality.
- How much does Chatterbox cost for a team or in production?
- Free weights; the cost is compute, and Nano runs on CPU. Artificial Analysis lists hosted Chatterbox at $25 and Chatterbox HD at $40 per million characters, against $0.65 for hosted Kokoro.
- Is Chatterbox safe for company data?
- Self-hosted, so text and reference audio stay local. MIT licence with no revenue limits; the only usage policy is 'Don't use this model to do bad things', and there is no consent check on cloned voices.
- What are the best Chatterbox alternatives?
- The closest alternatives we have reviewed are Kokoro-82M, Qwen3-TTS and Eleven v4 and v4 Turbo.
Answers checked Oct 22, 2026.