Voice quality in TTS used to be the thing that came up in every conversation, lately it doesn't, the bar has moved to latency. The questions are whether the model is light enough to run on the hardware it actually needs to live on without a cloud dependency propping it up, and what happens when you need thousands of concurrent voice agents and your autoregressive model is the bottleneck.
Chatterbox Nano and Chatterbox Flash are both answers to that. Both are open source, shipping today alongside DramaBox and Chatterbox Multilingual V3, which we released recently. One thing that holds across all of the releases: every model ships with watermarking by default. Whatever you build with them, the provenance is already there.
Chatterbox crossed 13 million downloads on Hugging Face somewhere in the middle of building all of this. It's a good reminder of how much this community actually builds with these models.
Chatterbox Nano
Chatterbox Nano is a 110M parameter TTS model built for the edge. It's small enough to run locally and fast enough that the gap between text and audio disappears.
- 10x realtime on GPU, 3x realtime on CPU using 8 threads
- Paralinguistic tags like [laugh], [sigh], [chuckle], and [cough] work natively
- Voice cloning from a 5-second reference clip, no fine-tuning required
- Watermarking baked in by default
It was trained on the same dataset as Chatterbox Turbo, so it carries the same expressiveness. The provenance signal is in the file the moment it is created. And like everything we are shipping this month, watermarking is baked in by default.
Welcome to Support Services. For account assistance, press 1. For billing questions, press 2. For technical support, press 3. To speak with a representative, press 0. To repeat these options, press star.
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. Don’t worry, we’ll sort this out together. Go ahead and tell me what you’re experiencing, and I'll walk you through the fix step by step.
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.
Chatterbox Flash
Chatterbox Flash is Chatterbox rebuilt on a diffusion-LLM architecture for teams running TTS at scale. Most production TTS today hits a ceiling on throughput that hardware alone cannot fix: autoregressive architectures generate tokens sequentially, you can add more machines but you cannot change the loop. Flash changes the loop itself.
- 2x the speed of our autoregressive baseline on vLLM
- Novel prior-subtraction technique that reduces diffusion steps without hurting output quality
- One of the first TTS models built on dLLM architecture that is actually ready to run in production
We think the prior-subtraction technique generalizes to dLLM TTS broadly, and it's a big part of why this is a production model rather than a research one. If throughput is the thing you are working around right now, this is worth a close look.
What and who these are for
Chatterbox Nano is for builders who need TTS to run where a server cannot follow:
- Voice agents at the edge
- On-device deployments
- Real-time applications where every millisecond is load-bearing
Chatterbox Flash is for teams running inference at volume who have hit the ceiling on autoregressive generation and need a way through it.
Both ship as open source under the MIT license.
How to get started
Model cards and quickstart codes for both are on Hugging Face - Chatterbox Nano and Chatterbox Flash.
.avif)
