#Kyutai TTS: A non-profit lab open-sources two voice models

A non-profit AI lab in France has open sourced two text-to-speech models: one with 100 million parameters that runs on a CPU, and a 1.6 billion parameter streaming model built for voice assistants.

#What it is

Kyutai is a French non-profit AI lab funded by Iliad Group, CMA CGM Group, and Schmidt Sciences. They previously built Moshi, a real-time conversational model, and they commit to open science — every model they release is open source. Their two TTS releases serve very different use cases.

Pocket TTS, released in January 2026, has 100 million parameters. It runs in real time on CPU with around 200ms latency, hits roughly 6x real-time on a MacBook Air M4, and uses only 2 CPU cores. It supports voice cloning from any audio file, handles unlimited-length text, and can run in the browser via WebAssembly. It covers six languages: English, French, German, Spanish, Portuguese, and Italian.

TTS 1.6B, released in July 2025, is a 1.6 billion parameter streaming model based on delayed streams modeling. It starts generating audio before it has received the full text, which makes it ideal for voice assistants where latency matters. It powers Unmute, a real-time voice AI system, and offers expressive voices — angry, calming, confused, sad, sarcastic, whisper, and even a jazz radio host.

#Why it matters

• Pocket TTS is small enough to run locally on a laptop CPU, in a browser, or embedded in an app — no GPU, no API key, no server. That is rare for voice cloning.
• Voice cloning from a single audio sample, across six languages, is genuinely useful for prototyping and accessibility — provided you have the voice owner's consent.
• The browser WebAssembly path and community ports to Rust, C++, C#, ONNX, Unity, and ComfyUI mean the model is already escaping into the wider ecosystem.
• TTS 1.6B's streaming architecture cuts latency for voice assistants, which is the metric that actually determines whether a voice bot feels responsive.
• Both models are open source under MIT and Apache licenses, with weights on Hugging Face.

#How it works

Pocket TTS installs with a single pip install pocket-tts. You can generate a WAV from the command line, serve a web interface on localhost, or call it directly from Python. Because the model is so small, it can run entirely in the browser via WebAssembly with no server behind it. Voice cloning works by feeding the model a reference audio file, after which it reads new text in a similar voice.

TTS 1.6B uses delayed streams modeling: it begins producing audio before the full text has arrived, which is the key trick for low-latency voice assistants. It requires a GPU server, in contrast to Pocket TTS's CPU focus. The two models are deliberately aimed at different deployment targets — local and embedded versus server and streaming.

#Caveats

Voice cloning should only be used with the consent of the voice owner. Pocket TTS has no GPU speedup because it uses batch size 1 and the model is very small — that is a design choice, not a bug. Pocket TTS also does not yet support inserting silence markers in text. TTS 1.6B needs a GPU server, so it is not a drop-in for the local-first crowd. The two models are not interchangeable; they serve different use cases.

#Who it's for

Pocket TTS is for anyone who wants to run TTS locally, embed it in an app, run it in a browser, or clone a voice on CPU. TTS 1.6B is for teams building voice assistants on a server who need streaming and low latency.

Two open source models, two deployment targets, one non-profit lab — pick the one that matches your stack.

Source: https://github.com/kyutai-labs/pocket-tts