Google DeepMind brings sign-language-to-text into Gboard

On 12 August 2026, Google DeepMind released a Sign-Language-to-Text (SL2T) model that translates sign language into text right inside the Gboard keyboard and the Live Transcribe app. Deaf and hard-of-hearing users can sign instead of type, and the translation appears immediately as text.

Key takeaways

The first rollout covers American Sign Language (ASL) translated to English; more languages are to follow.
The model was trained on more than 100,000 hours of footage across 50+ sign languages (about 25% ASL).
Instead of raw video, only the coordinates of body and hand landmarks are sent to the server.
The model reaches a zero-shot BLEURT score of 70 on the FLEURS-ASL benchmark.
The feature launches on the Pixel 11 at no extra cost, with more devices promised.

How the translation works

Instead of analyzing the video image, Google DeepMind uses MediaPipe Holistic to track body and hand landmarks. The raw video is discarded immediately, and only geometric coordinates are sent to the server. The model translates those directly into text, skipping the intermediate step of so-called glosses — simplified written labels for the meaning of individual signs.

100,000+hours of training data across 50+ sign languages

That is a meaningful break from older sign-recognition systems. The traditional chain ran from gesture through glosses to text, losing fluency and context along the way. Going straight from landmarks to text simplifies the pipeline and shortens the path to a natural sentence.

Privacy and quality

Sending only coordinates, rather than a recording of the face and surroundings, sharply limits the privacy footprint. Quality is measured with the BLEURT metric: with no prior tuning (zero-shot), the model scored 70 on the FLEURS-ASL benchmark. That is promising for translation without dataset-specific training, though everyday conversations will be the real test of usefulness.

Co-created with Deaf people

The project was built with the Deaf community. The company set up an AI Sign Language Advisory Committee (AISLAC) bringing together global Deaf organizations, and it was conceived by Sam Sepah, a Deaf Googler.

Signing in ASL is faster, more natural, and more delightful than typing in English.

Sam Sepah, project conceptualizer, Google.

Why it matters

Sign-language recognition long stayed a research niche that rarely reached users' hands. Wiring translation into the system keyboard turns it from a demo into an everyday accessibility tool. The landmark-based approach also shows that usefulness and privacy can coexist, without sending video recordings to a server.

What's next

After ASL and English, more sign languages are set to follow — the model was already trained on over 50 of them.
Beyond the Pixel 11, the feature is meant to reach more devices — Google promises expanded support.
The zero-shot BLEURT score of 70 is a baseline — future model versions can be compared on the same FLEURS-ASL benchmark.