Sign Language AI · Pixel 11
A phone that reads sign language
and turns it into text, on the spot.
Google DeepMind's sign-language model "SL2T" shipped inside the Pixel 11, unveiled on August 13. Sign in front of the camera and American Sign Language (ASL) becomes English text in real time — sign-language recognition moves from a research result to a standard feature on a mass-market phone for the first time.
From a research paper
to a feature you use daily
Sign-language recognition had stayed a Google DeepMind research paper and demo until the "Made by Google" event on August 12–13, when Pixel 11 launched with DeepMind's sign-language model "SL2T" (Sign Language to Text) built into Gboard and Live Transcribe as a standard feature. The feature converts American Sign Language (ASL) into English text in real time, and DeepMind laid out the technical details itself in an official blog post, "Putting sign language AI into users' hands."
Pixel 11 goes on sale August 20, and SL2T reaches everyday users alongside it. DeepMind's official account describes it as "our breakthrough sign language-to-text model powering new features for Deaf and hard of hearing users on Android" — from web search to composing a message to querying Gemini, any text field can now be filled by signing instead of typing. What makes this news is not another research demo, but a feature shipping on mass-market hardware.
No video leaves the phone —
only skeletal coordinates do
For privacy, SL2T never handles the raw camera feed itself. It works only with a sequence of points describing how the body moves.
On-device landmark extraction
A model built on Google's pose-estimation framework "MediaPipe Holistic" runs on the phone itself, converting the camera feed into a sequence of roughly 130 coordinates tracing the face, body and hands. The original video is discarded at this step and never sent anywhere.
Only coordinates reach SL2T
What reaches the server is numeric coordinate data, not video. Training data spans over 50 sign languages and more than 100,000 hours, with ASL reportedly making up roughly a quarter of the total.
Straight to text, no gloss step
Conventional sign-language translation often routes through an intermediate representation called "gloss," which tends to lose non-manual elements like facial expression and use of space. SL2T translates the coordinate sequence directly into English text, removing that bottleneck.
Signing instead of typing,
right where you already type
The feature isn't a standalone app — it's built into Gboard, an input surface people already use every day.
Typing, via Gboard
Set the keyboard language to en-US or en-CA and a camera icon opens sign-to-text input. It works in any text field — search boxes, messaging apps, wherever you'd normally type.
Face-to-face, via Live Transcribe
The existing feature that captions a hearing person's speech now gains a second path that turns a Deaf signer's side of the exchange into text too, making the conversation two-way on the spot.
Querying Gemini by signing
Instead of speaking to a voice assistant, users can sign questions and instructions directly to Gemini — a working example of AI assistant input that doesn't depend on speech.
Not built for Deaf people —
built with them.
Who this helps, and how
SL2T itself is reported to have originated with Sam Sepah, a Deaf Googler, and to have been developed together with an "AI Sign Language Advisory Committee" that DeepMind assembled from Deaf organizations and subject-matter experts. Deaf community members are said to have been involved from data collection through user evaluation and impact assessment, with a joint impact report reportedly published alongside this 1.0 release. That co-design process is itself a reference case for how other accessibility features might get built.
For designers, the lesson is in where the feature lives: rather than a new standalone app, it was slotted into Gboard and Live Transcribe — input surfaces people already use — instead of asking users to learn a new experience. For engineers, the architecture is a reusable pattern: extracting roughly 130 landmark coordinates on-device with MediaPipe Holistic and sending only that numeric sequence to the cloud model keeps gesture and pose recognition privacy-preserving while still supporting cloud-scale inference. For product managers, the rollout itself is instructive — starting with one language, ASL, rather than every sign language at once, and letting a Deaf advisory committee help set priorities is a workable template for sequencing an accessibility roadmap.
This isn't a story
of pure upside
An estimated 70 million people worldwide use sign language, yet SL2T currently supports only one direction — ASL to English. DeepMind describes expansion to other sign languages, and even sign generation (translating text back into signing), as future work with no stated timeline. Reporting suggests the model still stumbles on rare signs and rapid fingerspelling, and word-error rates broken down by signer demographics — beyond some notes on left-handed and one-handed signing — have not been disclosed.
Availability is also currently limited to a single device, the Pixel 11, with expansion to other Android phones described only as "coming soon." The real measure of progress for sign-language users isn't one feature on one phone — it's sign languages reaching the same level of support across digital products that spoken and written languages already have. Whether SL2T's ASL results hold up as it expands to dozens of sign languages, each with its own grammar and regional variation, will decide whether this becomes a lasting shift or a one-time headline.