The model watches body points, not the video, and turns American Sign Language into text on the device itself. The frame is discarded immediately.
In short
- On 12 August DeepMind announced the SL2T model inside Gboard and Live Transcribe on Pixel 11.
- It was trained on over 100,000 hours of material across more than 50 sign languages.
- American Sign Language to English goes first; others follow.
What stopped me first was the design, not the news itself. The model does not watch your video, it tracks where the joints of your hands and body are.
The facts: on 12 August 2026 Google DeepMind announced SL2T, a sign language to text translation model powering new features in Gboard and Live Transcribe on Pixel 11. Processing runs through an on-device model that tracks body landmarks rather than the frame itself; the video is discarded immediately. Training used over 100,000 hours of material across more than 50 sign languages. By Google's numbers the model reaches 70 BLEURT on the FLEURS-ASL benchmark with no prior tuning. American Sign Language to English comes first.
The small technical move here is also the privacy decision. Once coordinates go into processing instead of a face, the whole argument about what happens to the frame dissolves. Google says the video itself is discarded as soon as the points are read.
It works on one phone, for one language, in two apps. That is all for now.
Accessibility is measured in small things. A keyboard is exactly that.
The visual is generated code art. No third-party images.