Google DeepMind’s SL2T Brings Sign Language AI to Smartphones

For roughly 15 years, hearing smartphone users have been able to speak instead of type. On Aug. 12, 2026, Google DeepMind announced that Deaf and hard-of-hearing users will finally have an equivalent: a sign-language-to-text model called SL2T, now embedded in Gboard and Live Transcribe on the Pixel 11, starting with American Sign Language to English. More devices and additional sign languages are planned.

The practical experience mirrors voice dictation closely. Users can sign into any text field — drafting a message, running a web search, or prompting Gemini — without switching to a keyboard. In Live Transcribe, which already captions spoken audio in real time, signers can now sign their side of a conversation with a hearing person rather than typing it out. According to the team’s user testers, signing in ASL is faster and more natural than typing in English.

What Makes SL2T Different

The model’s architecture represents a meaningful departure from how most prior sign language AI systems were built. Earlier approaches typically relied on intermediate annotations called glosses — shorthand transcriptions of individual signs — as a stepping stone between video input and text output. SL2T skips that layer entirely, translating directly from body pose coordinates to text. The team argues this removes an artificial vocabulary ceiling and lets translation quality improve as training data grows.

That data pool is large: over 100,000 hours across more than 50 sign languages, with roughly a quarter in ASL. Training jointly on diverse languages and proficiency levels — rather than building isolated single-language models — appears to help the model learn shared underlying structures, outperforming single-language approaches in internal experiments.

On the FLEURS-ASL benchmark, SL2T achieves a zero-shot score of 70 BLEURT. To put that in context, academic baselines on the same benchmark had reached only 37.2 BLEURT, and leading general-purpose multimodal models — including large frontier systems — had shown virtually no capability on ASL translation tasks. The gap is categorical, not incremental.

Privacy was also a design constraint. Rather than sending raw video to a server, an on-device model called MediaPipe Holistic tracks the locations of body landmarks — points on the hands, arms, face and torso — and transmits only those geometric coordinates. The original video is discarded immediately on-device. The translation layer then operates on that coordinate sequence, not on footage of the user.

Built With, Not Just For

The project originated with Sam Sepah, a Deaf Googler, and Deaf collaborators were involved at every stage: data collection, user studies and impact evaluation. To formalize that involvement going forward, Google established the AI Sign Language Advisory Committee (AISLAC), which includes global Deaf organizations and subject-matter experts. The team also co-authored a public impact report covering SL2T 1.0’s capabilities and known limitations — an accountability mechanism that goes beyond a typical product launch.

Known limitations are disclosed honestly: the model can stumble on rare signs, rapid fingerspelling, passive grammatical constructions, classifier depictions and tense inference without sufficient context. These aren’t minor edge cases, but the transparency is notable.

“AI’s ability to process spoken languages has advanced rapidly over recent decades, enabling automatic translation, dictation, and conversational interfaces that feel effortless to hearing users. Yet this technological revolution has not reached the world’s more than 200 sign languages — and the estimated 70 million Deaf and hard of hearing people who use them,” Google DeepMind Sign Language Team wrote.

Competitive Landscape

The commercial sign language AI space has been active but pre-consumer. Companies like Signapse offer text-to-sign video output for BSL and ASL, and CirrusTranslate uses a hybrid AI-plus-human model for ASL translations. What neither has done is ship sign-to-text dictation inside a mobile keyboard at global scale. Gboard is one of the most widely installed mobile keyboards in the world; embedding SL2T there is a distribution advantage no sign language startup can match. Google itself announced a related model called SignGemma more than a year before this launch but acknowledged it hadn’t yet cleared internal quality thresholds — SL2T appears to be the deployment-ready result of that longer development arc.

Why This Matters to Students and Builders

For Deaf students, the immediate impact is straightforward: the same fluid, hands-free input that hearing classmates have used for a decade — drafting emails, querying AI assistants, searching while walking — is now accessible through sign language, at no additional cost on Pixel 11 hardware.

For students interested in building AI, the architecture is worth studying. The on-device tracking layer relies on MediaPipe, an open-source framework from Google with Python and JavaScript APIs, meaning the foundational perception stack that SL2T sits on top of is already publicly available to experiment with. The proprietary piece is the translation model above it — and Google has indicated that API access and potential open-weights releases under the SignGemma name are under active discussion, which would open the door for Deaf researchers, developers and Deaf-owned businesses to build on the architecture.

For students focused on responsible AI, AISLAC and the co-authored impact report offer a concrete, citable example of participatory governance in action — the kind of model that increasingly comes up in ethics coursework and job interviews at AI labs.

SL2T is available now in Gboard and Live Transcribe on the Pixel 11 at no additional cost.

Source: Google DeepMind