Google debuts SL2T, an AI model that’s designed to understand sign language

Google DeepMind is trying to do for sign language what phones already do for speech: let people talk to devices instead of typing everything out. Its new model, SL2T, debuts on the Pixel 11 and powers a sign-to-text feature in Gboard and Live Transcribe. For now, it understands American Sign Language and turns it into English text.

That matters because speech recognition has raced ahead while sign-language support has mostly lagged behind. Google says there are about 70 million deaf or hard-of-hearing people worldwide, and that the AI industry has largely left them out. SL2T is meant to change that by letting users sign into their phone’s camera to search the web, draft emails and text messages, edit documents, and even prompt Gemini.

The model was trained on more than 100,000 hours of sign-language data covering over 50 languages, with roughly a quarter of that coming from ASL. Google says it used that broader mix to learn patterns shared across sign languages, even though the first release only handles ASL. The company also says SL2T is the first model of its kind to ship inside a real consumer product.

Google is also leaning hard on privacy. SL2T uses an on-device vision system called MediaPipe Holistic to track the position of the signer’s face, hands, arms and torso, then sends coordinates to the cloud instead of raw video. That setup is supposed to make it faster and more secure. The model also skips glosses, which Google says helps it capture ASL’s facial expression and spatial grammar, and it includes support for left-handed signers and one-handed signing.

The company says the system reached a score of 70 BLEURT on the FLEURS-ASL benchmark. And it’s not stopping there: Google has set up an AI Sign Language Advisory Committee with deaf organizations and sign-language experts, and it plans to expand SL2T to more sign languages and eventually build sign-language generation models.