Google DeepMind brings multilingual sign-language-to-text AI to Pixel 11

Sign-language AI moves from research into a consumer product

Google DeepMind announced on August 12, 2026 that its new sign-language-to-text model, SL2T, is powering sign-to-text features in Gboard and Live Transcribe on Pixel 11. The initial product experience translates American Sign Language into English, with DeepMind saying additional devices and languages are planned. That makes the release notable not simply as a benchmark result, but as a deployment aimed at everyday communication for Deaf and hard-of-hearing users.

A multilingual training approach

DeepMind says SL2T was trained on more than 100,000 hours of data spanning over 50 sign languages, with roughly one quarter of the training data in ASL. The model translates visual signing directly into text rather than relying on an intermediate gloss representation. The company argues this direct approach can better preserve aspects of sign languages that do not map cleanly to a simple sequence of spoken-language words.

The system must solve both visual-perception and language-translation problems. Sign languages carry meaning through hands, arms, facial expression, head position and body movement, while also having their own grammar and vocabulary. DeepMind reports that SL2T reached a zero-shot score of 70 BLEURT on the FLEURS-ASL sd-test benchmark, while also acknowledging that real-world usability requires work beyond benchmark performance.

Privacy and on-device preprocessing

A key implementation detail is the treatment of camera data. DeepMind says MediaPipe Holistic runs on device to convert video into pose-landmark coordinates. The server receives those geometric coordinates for translation, while the original video is discarded immediately. This architecture is intended to reduce the amount of raw visual data that leaves the device while still enabling server-side translation.

Product use and remaining limitations

In Gboard, users can sign where they would normally type, such as when drafting a message, searching the web or interacting with Gemini. In Live Transcribe, signing can be used for responses during conversations. DeepMind also describes work on streaming latency, hallucinations when no signing is present, left-handed signing and one-handed signing while a user holds a phone.

The release is not a claim that sign-language translation is solved. DeepMind's own examples show occasional errors on rare signs, rapid fingerspelling, tense and classifier constructions. The company says it worked with Deaf users and established an AI Sign Language Advisory Committee to include community perspectives in development and deployment decisions.

For developers and AI researchers, the broader significance is the combination of multilingual visual-language modeling, privacy-oriented preprocessing and real consumer deployment. For users, the practical value will depend on translation quality across varied signing styles, contexts and future languages as the system expands beyond its initial ASL-to-English release.