Google DeepMind brings sign language translation to the phone

The SL2T model, trained on more than 50 sign languages and 100,000 hours of data, arrives in Gboard and Live Transcribe on Pixel 11.

Google DeepMind launched its SL2T (Sign-Language-to-Text) model on August 12, 2026, which translates sign language directly into text. The first open pair is the American Sign Language to English, but the model was trained on over fifty sign languages and over 100,000 hours of data; others will follow.

On the FLEURS-ASL test bench, the system achieves a BLEURT score of 70 in zero-shot configuration, well above any published result. The technical difference lies in the abandonment of the "gloses", these word-to-word intermediate representations which traditionally served as pivot: SL2T translates directly from the characteristic points of the posture, extracted on the device via MediaPipe Holistic. The choice also protects privacy, as raw video does not leave the phone. The model manages the sign with one hand like left-handed signers.

The feature is available in Gboard and Live Transcribe on Pixel 11, with an announced deployment on other devices. The project was designed by Sam Sepah, a deaf engineer at Google, and developed with an advisory committee of organizations of deaf people associated with system design, evaluation and governance.