Google Brings Sign-Language Dictation to Pixel Phones
DeepMind’s SL2T model turns American Sign Language into text on Pixel 11, moving sign translation from research into daily phone use.
From benchmark to input method
Google DeepMind has introduced SL2T, a multilingual sign-language-to-text model that now powers sign dictation in Gboard and two-way use in Live Transcribe on Pixel 11 phones. The initial product translates American Sign Language into English, allowing users to sign when composing messages, searching the web or issuing requests to Gemini. Google says more devices and language pairs are planned.
Instead of uploading raw camera footage, the system first runs an on-device perception model that extracts movement coordinates from the user’s face, hands, arms and torso. Those coordinates are sent for translation, reducing the amount of identifiable visual information that leaves the phone. The model is designed to handle practical conditions that conventional sign-language datasets often miss, including one-handed signing while the other hand holds a phone, left-handed signing and incidental movements that should not be interpreted as language.
DeepMind says SL2T was developed with Deaf employees and an external sign-language advisory committee. Its training work spans more than 50 sign languages, although only ASL-to-English is shipping at launch. That distinction is important: sign languages are independent natural languages with their own grammar, not gesture-based encodings of surrounding spoken languages, so expansion requires more than substituting a new vocabulary.
Why it matters
Sign-language recognition has produced years of research demonstrations without becoming a dependable, general-purpose input method. Integrating SL2T into a keyboard and a live conversation tool changes the test from benchmark accuracy to whether Deaf users can rely on it throughout the day. The launch is narrow—one phone generation and one translation direction—but it establishes an unusually concrete consumer use for multimodal AI. It also offers a privacy-conscious architecture that other camera-based assistants may be expected to match.