DeepMind ships a sign-language-to-text model

Google DeepMind released SL2T (sign-language-to-text) on August 12, a multilingual model that converts sign language into written text for dictation, search, messaging, and document composition, the sign-language equivalent of voice dictation for hearing users. Rather than processing raw camera video, SL2T works from a sequence of pose landmark locations, a design choice DeepMind frames as privacy-protective since no video of the signer needs to leave the device or get stored.

DeepMind trained the model on upward of 100,000 hours of footage spanning over 50 different sign languages, roughly a quarter of it American Sign Language, and DeepMind reports a zero-shot score of 70 BLEURT on the FLEURS-ASL benchmark, which it says is significantly higher than any previously reported score for sign language translation. SL2T is live now on Google Pixel 11 through Gboard and Live Transcribe, starting with ASL-to-English, with more devices and languages coming.

What it means for you

The interesting engineering decision here isn’t the translation accuracy, it’s building the pipeline on pose landmarks instead of raw video, which cuts both the privacy exposure and the compute cost of running a vision model on every frame. If you’re building any kind of gesture, sign, or motion-based input into a product, this is the reference pattern to look at before you default to streaming raw camera data through a model, the same instinct behind setting an honest data-privacy baseline before you assume you know what a platform is actually capturing. It’s the same week DeepMind shipped Gemini 3.7 Flash, a reminder that a lab’s research releases and its flagship model line move on separate, overlapping schedules worth tracking independently.