Google brings ASL-to-text dictation to Pixel 11

Signing becomes another way to type

For many Deaf people, entering text on a phone still means typing in English, even when American Sign Language is the language they use most naturally. Google is adding another route on Pixel 11: sign to the camera and receive an English draft where text normally goes.

The new sign-language-to-text model, called SL2T, launches in Gboard and Live Transcribe. In Gboard, a person can sign a search, message, note or Gemini request. Live Transcribe uses the same model for short, one-to-one conversations.

The important word is draft. The output appears for the signer to review and edit before it is sent, shown or spoken. A joint impact report co-authored by Google DeepMind and members of its AI Sign Language Advisory Committee describes the model as an assistive input tool for low-stakes use, not an autonomous interpreter.

That makes this a meaningful product release. It does not make it a general solution for sign-language access.

The phone sends landmarks, not video

The camera feed is processed on the device with MediaPipe Holistic. It extracts 130 two-dimensional points that track the face, body and hands frame by frame. Google says the raw video is then discarded, while the geometric coordinates are sent to its servers for translation.

According to the impact report, Google does not retain logs of user inputs or model outputs unless a user has explicitly agreed to take part in an evaluation study. That is the company's stated data-handling boundary for this release.

Landmarks remove a lot of visual detail, including a person's appearance and much of their surroundings. They also remove useful information. The representation does not track the tongue, has no depth ordering and cannot see the real-world objects a signer may be referring to.

SL2T translates those landmarks directly into English rather than first turning them into a written list of signs. Google says the model was trained on more than 100,000 hours across over 50 sign languages, with about a quarter of the data in ASL. The product shipping now supports only ASL to English.

A benchmark leap is not the whole product

Google reports a zero-shot score of 70 BLEURT on the FLEURS-ASL sd-test set, above the previous result it cites. The impact report lists additional results for one-handed signing, assistant-style phrases and fingerspelling.

Those are company-reported measurements. They help describe progress, but they are not an independent comparison of the finished feature against every alternative, and a benchmark cannot reproduce the full range of everyday signing.

The impact report is more revealing when it names what still goes wrong. The model can miss facial grammar such as eyebrow movement or a head shake. It can confuse regional or polysemous signs, struggle in poor light, guess at proper nouns and produce text when a person pauses or another person enters the frame.

Each clip is capped at 60 seconds and translated without memory of earlier turns. The signer also needs enough English literacy and device confidence to spot and fix a bad translation. Human review is not an optional extra here. It is part of the design.

Community testing found utility and friction

The advisory group included representatives from the National Association of the Deaf, RIT's National Technical Institute for the Deaf, the Deaf Professional Arts Network and the World Federation of the Deaf, along with an independent subject-matter expert.

Advisors tested pre-release software through daily use and structured tasks. They reported strong utility for short searches, messages, notes and casual exchanges. Some said signing into documents reduced typing fatigue. They also found workflow freezes, awkward mode switching, missing editing commands and inconsistent recognition of regional language and non-manual markers.

The report draws a firm line around high-stakes use. Medical care, legal proceedings, classroom instruction, employment decisions and government services are all outside the intended scope. The system does not translate spoken language back into sign, cannot ask for clarification and does not carry the judgement or accountability of a qualified interpreter.

That boundary matters because a cheap software option can be tempting to an institution. Google and the advisory group say SL2T should not be used to avoid legal accessibility duties or replace human interpreters where they are required.

What is confirmed, claimed and still open

Confirmed: Google DeepMind announced SL2T on 12 August 2026 and says the first release brings ASL-to-English input to Gboard and Live Transcribe on Pixel 11 in the United States. The accompanying impact report was co-authored with named Deaf organisations and experts.

Google's claims: the training scale, benchmark scores, server-side log policy and improvements made between the July test build and the public release all come from Google and the joint report. They have not been independently reproduced in this product setting.

The documented limits: children were not included in training or formal evaluation, testing across signers with motor disabilities remains limited, and demographic breakdowns for regional varieties and Black ASL are still work in progress. Multi-signer scenes and personalised sign lists are also absent.

Still open: how well the launch version works across the wider Deaf community, when more devices and sign languages will arrive, and whether Google's privacy and quality claims will receive independent scrutiny.

For now, the clearest value is modest and practical: another way to put a short thought into a phone, with the signer still firmly in charge of the final words.