chatterbox nano • resemble TTS research

Fast, local TTS for
real-time voice agents.

Get started
10x
Real-time on GPU
3x
Real-time on CPU
(8 threads)
110M
Parameters
Trusted by
the problem

Real-time agents can't wait for full-scale models.

Most TTS models are built for cloud rendering — large, GPU-hungry, and dependent on a network round-trip. Voice agents, local applications, and mobile deployments need speech that runs fast without that dependency. Until now, that meant choosing speed or quality. Chatterbox Nano removes that trade-off.
WHAT IT DOES

Built for deployments where speed is the constraint.

Nano doesn't cut corners to get small. It brings the full Chatterbox feature set to environments that couldn't run it before.

Edge-ready architecture
110M parameters run locally with minimal VRAM — suited for on-device, mobile, and air-gapped deployments.
10x real-time on GPU
Fast enough for natural turn-taking in real-time voice agents. 3× real-time on CPU using 8 threads.
Zero-shot voice cloning
Clone any voice from 10 seconds or more of reference audio. No fine-tuning. Ready for synthesis immediately.
Native paralinguistic tags
Trained on the same dataset as Chatterbox Turbo. Supports [laugh], [sigh], [cough], [chuckle], [gasp], and more.
PerTh watermarking
Every clip is watermarked at generation — imperceptible to listeners, persistent through re-encoding, verifiable on demand.
MIT licensed
Self-host via pip, deploy on-prem, or run air-gapped. Weights on GitHub and Hugging Face. No commercial restrictions.
Listen

Hear Chatterbox Nano

Zero-shot voice cloning from short reference audio. Paralinguistic tags rendered directly in the cloned voice.

IVR • SUPPORT AGENT
"Good afternoon, thanks for calling IT support. I'm Alex, your support agent [chuckle]. Let's take a look at what's going on with your account."
HEALTHCARE • INTAKE AGENT
"Thank you for calling CareLine Medical. I'm Howard, your support agent. Please tell me your name and date of birth so I can get your file."
how the model works

Same capability set as Turbo. A fraction of the size.

Nano is trained on the same dataset as Chatterbox Turbo — full expressive range, all paralinguistic tags, and zero-shot voice cloning included. The difference is parameter count and where it can run. At 110M parameters, Nano fits where Turbo can't: edge devices, mobile deployments, on-prem servers with limited GPU.
deployment

Run it where your infrastructure requires.

Cloud API

Resemble managed platform. RESTful API with streaming. Business plan or higher for Voice Cloning API.

MANAGED
Self-hosted

Install via pip. Weights on Hugging Face. Run on your own GPU infrastructure.

PIP INSTALL RESEMBLEAI
On-prem / Edge

Docker and Kubernetes for data sovereignty requirements. 3× real-time on CPU.

DOCKER • K8S • EDGE
Frequently asked questions
What is Chatterbox Nano?
Chatterbox Nano is Resemble AI's smallest and fastest open-source TTS model. At 110M parameters, it runs 10× real-time on GPU and 3× real-time on CPU, making it suited for edge deployments, mobile applications, and real-time voice agents that can't rely on a cloud round-trip.
How does Nano compare to Chatterbox Turbo?
Nano is trained on the same dataset as Chatterbox Turbo, so it shares the same expressive range, paralinguistic tag support, and zero-shot voice cloning. The difference is size — at 110M parameters versus Turbo's 350M, Nano is designed for environments where Turbo would be too resource-intensive.
Does Nano support voice cloning?
Yes. Nano supports zero-shot voice cloning from 10 seconds or more of reference audio. No fine-tuning or training run is required — provide a reference clip and the cloned voice is immediately available for synthesis.
What paralinguistic tags does Nano support?
Nano supports the same paralinguistic tags as Chatterbox Turbo, including [laugh], [sigh], [cough], [chuckle], [gasp], and more. Tags are rendered directly in the cloned voice with no post-processing required.
Can I run Nano on-prem without internet access?
Yes. Nano is MIT-licensed and available on GitHub and Hugging Face. Install via pip and run on your own GPU or CPU infrastructure. No connection to Resemble's managed platform is required.
What is PerTh watermarking?
PerTh is Resemble AI's neural speech watermarker. Every clip Nano generates carries a PerTh watermark — embedded at the perceptual threshold so it is imperceptible to listeners, persistent through re-encoding and format conversion, and verifiable on demand to confirm the audio's origin.
Get complete generative AI security
Join thousands of developers and enterprises securing with Resemble AI
Contact us