Sign Language AI: SL2T Translates ASL on Pixel 11

Sign language AI now works in Gboard and Live Transcribe on Pixel 11

Google DeepMind's sign-language-to-text model, SL2T, now powers sign-to-text dictation on Pixel 11 in Gboard and Live Transcribe, turning ASL into English at no extra cost. The Google DeepMind Sign Language Team announced it in a post dated August 12, 2026. The team calls it the first time sign language AI has left the lab for consumer products.

Google's Pixel 11 how-to post followed on August 21, 2026, when the Pixel 11 series was already in stores. The DeepMind post presents the feature as dictation for signers: hearing users can speak instead of typing, and now Deaf users can sign to the phone wherever they would type.

Gboard lets you sign to search, draft messages, and ask Gemini

Since Gboard is the Pixel 11's keyboard, SL2T lets a Deaf user sign anywhere they would otherwise type. Google names three uses: searching the web, drafting messages or documents, and asking Gemini to answer questions or do tasks. Its testers, Google says, found signing in ASL faster, more natural and more delightful than typing in English.

Setup takes two steps, according to Google's Pixel 11 post: add sign-to-text to the Gboard toolbar, then sign in ASL to the front camera. The translation appears on the outer display. The feature reads hand shapes, facial expressions and body movements and writes the ASL out as English sentences in real time.

In Live Transcribe, you can sign replies instead of typing them

Live Transcribe covers conversations, letting a Deaf user sign replies instead of typing back and forth. The Pixel 11 post adds that both one-handed and two-handed signing work, so signers can use their native language in Live Transcribe or anywhere on the device through Gboard.

Neither Google post says what the hearing person in the conversation sees or does. Both cover only the Deaf user's side, where signed replies become text, and neither gives setup steps specific to Live Transcribe.

ASL to English on Pixel 11 is what ships at launch

At launch the feature is ASL to English on Pixel 11, one direction only, in Gboard and Live Transcribe. Google says more devices are coming soon and other languages will follow, but gives no list or dates. The table below sums up what ships.

SL2T at launch

	Status at launch
Direction	Sign to text only
Language pair	ASL to English
Apps	Gboard and Live Transcribe
Devices	Pixel 11 first; more devices "coming soon", with no date
Price	No additional cost
Other sign languages	Google says more will follow; no list or dates

This release works only from sign to text, so the phone does not turn English into signing. Google lists sign language generation among its next steps.

SL2T turns pose landmarks into English text, skipping the gloss step

SL2T converts pose landmarks, the points a phone tracks on a signer's body, directly into English text without going through glosses, the written labels for signs that earlier translation work relied on. Google says only the coordinates of those points reach the server, never the camera video. Translating straight from landmarks, it adds, removes artificial vocabulary limits and lets translation quality improve as data grows.

Sign language translation needs vision and real machine translation

Google describes sign language translation as two hard problems in one: real machine translation between separate languages, and computer vision that can follow movement across the whole body. Spoken-language AI has advanced quickly, Google notes, but that progress has not reached the more than 200 sign languages used by an estimated 70 million Deaf and hard of hearing people. For Deaf communities, sign languages are the primary languages and central to cultural identity.

The first problem is language. Speech-to-text maps sound to written words in the same language step by step. ASL is a natural language with its own grammar and vocabulary, so ASL to English is real translation between two languages, not a sign-by-sign swap into English words.

The second problem, vision, comes from how a sign language carries meaning: through the hands, arms, torso, head and face moving together. Tracking all of it at high frame rates is demanding computer vision work, and Google says early attempts such as sign language gloves fell short because sign languages are not simply "English on the hands." Reading them takes fine-grained perception of whole-body movement plus full translation, and SL2T is built for both.

Only pose landmarks, not video, reach the server

SL2T works from a sequence of pose landmarks instead of the raw camera feed, according to Google, and only their coordinates go to the server. An on-device model called MediaPipe Holistic tracks those points on the phone itself.

Google cites user privacy as the reason for working from landmarks. Since only coordinates are sent for translation, the original video can be thrown away immediately.

On the server, SL2T converts the coordinate sequence into streaming text without glosses. Much earlier research on sign language translation depended on glosses, which, Google says, miss the rich, non-linear side of signing, such as non-manual markers (signals made without the hands) and spatial constructions.

Training covers 50+ sign languages, with about a quarter in ASL

SL2T was trained on more than 100,000 hours of data from over 50 sign languages, about a quarter of it ASL. Google describes the approach as a user-centered, culturally informed design paired with large-scale data. The languages, dialects and skill levels are trained together rather than in separate models.

In Google's experiments, training them together taught the model structures common to sign languages, and it outperformed models trained on one language. Still, only ASL to English is turned on at launch, and the post does not say which of the other trained languages comes next.

Google reports 70 BLEURT zero-shot on FLEURS-ASL and names its errors

SL2T scores 70 BLEURT zero-shot on FLEURS-ASL (sd-test), Google reports, well above any score reported before. Google calls FLEURS-ASL a benchmark for ASL to English translation quality and SL2T the most capable sign language translation model so far on key benchmarks like it.

BLEURT is a learned metric that scores how close a translation comes to a reference, and zero-shot means the model was not tuned on the benchmark's data beforehand. The score is Google's own, and Google also lists the practical problems it tackled and the errors still left.

Google also worked on latency, hallucination, and left-handed signing

A strong benchmark score alone does not make a model usable in daily life, Google says, so the team also worked on four practical problems. It names all four in the same paragraph as the score:

• Streaming latency: keeping the wait short while text streams out during signing.
• Hallucination on non-signing inputs: stopping the model from producing text when what the camera sees is not signing.
• Fairness for left-handed signers, who make up 10% of signers according to Google.
• One-handed signing, for when the other hand is holding the phone.

None of the four comes with a measured result. The post gives no latency figure and no accuracy number for left-handed or one-handed signing.

ASL to English errors remain in fingerspelling and classifiers

The caption under Google's examples admits that occasional ASL to English errors remain and names five kinds. Google shows five FLEURS-ASL examples, each pairing the original English sentence with SL2T's output, and the named errors show up in only two of them.

Error types named in Google's caption

Error type	Example Google gives
Rare signs	None shown
Rapid fingerspelling	"prey" came out as "grey"
Passive constructions	None shown
Classifier depictions	"claws" was dropped
Tense without context	"kicked off" became "start"

Those errors come from two sentences. The original bird sentence describes a feathered bird of prey with claws like a Velociraptor's, believed to have walked on two legs. SL2T's output says the creature "eats grey" and drops the claws.

In the rugby sentence, the tense changes. The original says the games "kicked off" at 10:00am, while SL2T says games "start" at 10 a.m. The other three examples, on the Cook Islands, income tax in federal countries and great-grandchildren on an alien world, have no error named in the caption.

Rare signs and passive constructions are only named in the caption, with no example. Google gives no error rate or count per type, so the table shows what kinds of mistakes occur, not how often.

Deaf community input shaped Google's sign language AI

Deaf perspectives shaped every stage of SL2T, from first concept to impact assessment, Google says. The project was conceived by Sam Sepah, a Deaf Googler.

Deaf partners helped gather the data, Deaf user studies evaluated the model, and Deaf experts assessed its impact. Google sums up the principle as building with the Deaf community, not only for it.

The AISLAC advisory committee and a joint impact report for SL2T 1.0

Google set up the AI Sign Language Advisory Committee (AISLAC) and co-wrote a joint impact report for SL2T 1.0 in Gboard and Live Transcribe. The committee brings together many Deaf organizations from around the world and subject-matter experts. Google calls this participatory governance, where the communities most affected by a technology directly shape its development priorities.

According to Google, the report sets out the technology's capabilities and current limits. Google plans the same joint approach for every major sign language release. The post neither summarizes the report's findings nor links to it.

Additional sign languages and sign language generation are next

Google's next steps are more sign languages, sign language generation and what it calls frontier AI capabilities, none with dates. Today's release works one way, from signing to text. Google calls bringing ASL input to phones only the beginning.

Google states the goal behind that list plainly. It links the work to its mission of making the world's information universally accessible, and says universal accessibility means full parity with spoken and written languages.

SL2T was built jointly by teams from Google DeepMind and Android. Google's acknowledgements list 15 people on the core team that developed the model and 13 on the Android team that put it into both apps.

Google's post likens sign-to-text to voice dictation. For the speech side of that comparison, Gemini 3.5 Transcribe: pricing and API limits covers Google's speech-to-text API.

On Pixel 11, sign language AI now means ASL to English sign-to-text in Gboard and Live Transcribe, built on pose landmarks instead of video. Google scores it at 70 BLEURT zero-shot, lists the occasional errors that remain, and says more languages and devices will follow, without dates.