Google DeepMind expands AI, search access with sign-language-to-text launch

Google DeepMind has released a sign-language-to-text model called SL2T, and it will be inside Gboard and Live Transcribe on the new Pixel 11. 

The company says it is the first sign language AI to ship in a real consumer product. 

Signing to the phone instead of typing

People who can hear and speak have been able to use voice dictation to talk to their devices for years. 

Now, those who cannot weave a speech to utilize voice dictation but use sign language can now interact with the internet, draft a message, or edit a document using their Pixel 11 device, thanks to SL2T. 

The signers can also hand a task to Gemini as well 

In Live Transcribe, a user can sign a reply during a face-to-face conversation rather than tapping one out.

Currently, the model handles American Sign Language (ASL) to English only. Google says that the model will be rolled out to more devices in the future, along with more languages.

Model testers reportedly described signing in ASL as quicker and more natural than typing in English.

Why is sign language harder than speech for AI?

DeepMind sees sign languages as full natural languages, not “English on the hands.” 

However, its development has been lagging when compared to spoken-language AI, largely because of two challenges.

The first one is that sign languages come with their own grammar and vocabulary, so the system has to translate them rather than transcribe. They also convey meaning through the hands, arms, torso, head, and face at once, which is a demanding computer-vision problem to track accurately.

There have been earlier efforts in developing these solutions, but most of them failed. One of them, sign-language gloves, failed partly because they read hand shapes alone and ignored the rest of the body. SL2T was built to handle both the visual perception and the translation.

How the model handles video and privacy

SL2T does not send raw footage to the cloud. An on-device computer-vision component called MediaPipe Holistic maps the signer’s face, hands, arms, and torso into geometric coordinates, and only those coordinates travel to Google’s servers. DeepMind says that keeps the actual video on the phone and makes the pipeline faster.

The model translates those landmark sequences straight into text, skipping the intermediate labels known as glosses. According to DeepMind, glosses cap vocabulary and miss the non-manual markers and spatial grammar that ASL relies on. 

Google trained SL2T on more than 100,000 hours of data spanning over 50 sign languages, with about a quarter of it in ASL, and reported a score of 70 BLEURT on the FLEURS-ASL benchmark. The model also supports one-handed and left-handed signing, includes latency optimizations, and has mechanisms meant to stop it from hallucinating text when the camera catches non-signing movement.

What is Google building around the launch?

Google has set up an AI Sign Language Advisory Committee with Deaf organizations and sign-language experts to guide how the technology is deployed, and the committee will co-author a report on what the model can and cannot yet do. 

Next on the roadmap is the addition of more sign languages and models that generate sign language rather than only read it.

The consumer release has been building for a while. Android Authority spotted a sign-to-text option inside a Gboard beta on July 17, 2026, after DeepMind had earlier teased a sign-language model called SignGemma. 

The launch also lands during a turbulent stretch for the lab. In early August 2026, Demis Hassabis moved from DeepMind CEO to chairman, and Koray Kavukcuoglu took over day-to-day operations, with the Gemini app now past 950 million monthly users.

If you're reading this, you’re already ahead. Stay there with our newsletter.