Sam Witteveen · 2026-08-31 · notable
Sam Witteveen — 'BreezeTTS2: 100% Local Real-Time Voice'
Sam Witteveen walks through Breeze-TTS-2, a 3B open-weights text-to-speech model that runs entirely on local hardware in real time. The model card puts it at #1 among open-weight models on the Artificial Analysis TTS leaderboard.
A hands-on look at running a leaderboard-topping open-weights speech model on your own machine.
Key specs
| Parameters | 3B |
|---|---|
| Time to first audio | under 40 ms |
What is it?
Breeze-TTS-2 is the subject of this Sam Witteveen video: a 3-billion-parameter open-weights text-to-speech model that, in his words, "can run 100% locally in real time." BreezeBlue (RESONIA, INC.) released the weights and PyTorch inference code on August 25, 2026, and Witteveen published his walkthrough on August 31.
How does it work?
The model streams audio rather than rendering a whole clip first, which is what makes conversational use possible. Its card reports under 40 ms time to first audio and roughly 3.1x real-time generation on an NVIDIA H100 with the fast path enabled. Beyond plain reading, Breeze-TTS-2 supports voice cloning from a reference clip, voice design from a written description, and voice direction that steers tone while keeping the speaker's identity.
Why does it matter?
Good speech synthesis has mostly meant sending audio to a paid API. A 3B model that fits on one GPU changes the calculation for anyone building voice agents, accessibility tools, or offline apps. Witteveen's walkthrough matters because it shows the local setup working end to end, not just a leaderboard position — the model card ranks Breeze-TTS-2 first among open-weight systems on the Artificial Analysis TTS leaderboard.
Who is it for?
developers building voice agents and offline speech apps
Try it
https://huggingface.co/BreezeBlue/Breeze-TTS-2