Google SL2T Brings ASL Sign-to-Text to Pixel 11

Google DeepMind’s sign-language-to-text model, SL2T, is leaving the research-demo category and entering two ordinary Android surfaces. On Pixel 11, a person can sign in American Sign Language and receive English text inside Gboard or Live Transcribe. That turns signing into an input method anywhere Gboard can type, and lets someone answer a spoken conversation in Live Transcribe without switching to a keyboard.

This is not “English on the hands” recognition. ASL has its own grammar and conveys meaning through simultaneous hand, arm, torso, head and facial movement. SL2T performs visual language translation: an on-device model converts movement into geometric landmarks, then a server model translates the landmark sequence into streaming English.

TL;DR: what people are asking

Why sign language needs translation, not transcription

Spoken transcription usually maps an audio sequence into written words in the same language. Sign-language input is a different problem. ASL is not manually encoded English, and a sign cannot always be understood from hand shape alone.

Meaning can depend on:

• movement and orientation of both hands;
• where a sign occurs in three-dimensional space;
• torso and shoulder position;
• head movement and eye gaze;
• facial expression and other non-manual markers;
• spatial references established earlier in the conversation.

That is why glove-based systems and isolated-sign classifiers have a ceiling. A glove may capture finger flexion while missing the facial grammar or spatial construction that changes the sentence. SL2T instead treats the input as a coordinated, time-varying whole-body signal.

The same lesson appears across other accessibility interfaces. Google’s Magic Pointer interprets body movement as intentional pointer control, while the Augmental MouthPad and Neuralink wheelchair demo translate different physical signals into computer commands. The product succeeds only when the input representation matches how people actually communicate or act.

How SL2T works on Pixel 11

The runtime has two distinct privacy and compute zones:

Pixel camera
  → MediaPipe Holistic on-device
  → hand, face and body landmark coordinates
  → raw video discarded
  → coordinate sequence sent to Google servers
  → SL2T translation
  → streaming English text in Gboard or Live Transcribe

Google says SL2T bypasses glosses, hand-authored labels often used as an intermediate representation in sign-language research. Gloss vocabularies simplify evaluation, but they can flatten non-linear grammar and impose an artificial list of allowed signs. Direct landmark-to-text training lets the output vocabulary scale with the text model rather than a gloss dictionary.

The model was trained on more than 100,000 hours across over 50 sign languages, with roughly a quarter of that material in ASL. Google says multilingual training improved results compared with single-language systems because related visual and linguistic structures transfer across datasets, dialects and proficiency levels.

That does not make the consumer product multilingual on day one. The training system covers more than 50 languages; the shipping feature is ASL to English only. Conflating those two claims would turn a research direction into a false availability promise.

What the privacy design protects, and what it does not

Extracting landmarks locally is materially better than uploading continuous camera video. Google says the raw feed is discarded immediately, so servers do not need the signer’s room, clothing texture or full-resolution face to perform translation.

But the feature is not fully local. The landmark stream still represents a person’s body and facial movement over time, and translation happens in the cloud. A responsible privacy reading asks:

1. How long are landmark sequences, translations and diagnostic logs retained?
2. Are they used to improve models, and under which account or telemetry setting?
3. Can the feature operate when connectivity is absent or unreliable?
4. Can an enterprise, school or healthcare deployment control residency and logging?
5. How are bystanders and other signers handled when more than one person enters frame?

Google’s launch post explains the video-to-landmark boundary but does not answer every operational question above. “The video never leaves the phone” is useful and true according to the company; “all sign-language processing is on-device” would be false.

This split architecture resembles the general edge-AI pattern described in the Coral on-device AI guide: keep high-bandwidth or highly identifying perception near the sensor, then send a reduced representation to a more capable cloud model. It lowers exposure without eliminating cloud trust.

How good is the translation?

Google reports a zero-shot BLEURT score of 70 on the FLEURS-ASL sd-test split, which it says is substantially higher than previously reported results. BLEURT estimates semantic similarity between generated and reference text. It is a useful benchmark signal, not a guarantee that every translation preserves the signer’s meaning.

Google’s own launch table is more informative than the headline because it includes mistakes:

Google also optimized for product conditions a lab benchmark can miss: left-handed signers, one-handed signing while holding a phone, streaming latency, and not hallucinating text when the person is not signing. Those slices should be reported separately over time, especially across skin tones, camera angles, signing dialects, mobility differences and real lighting.

For anyone evaluating the claim, use the same discipline as reading an AI benchmark: name the split, metric, language direction, device path and error categories. “State of the art” does not describe performance on an individual signer.

What Pixel 11 users can do now

In Gboard, ASL input can generate text anywhere the keyboard works: search boxes, messages, documents and prompts to Gemini. That matters because accessibility becomes a system input method rather than a special app a user must convince everyone else to install.

In Live Transcribe, the flow becomes two-way. The phone can render a hearing person’s speech as text, while a Deaf user signs a response that SL2T converts into English text. Google does not say the release synthesizes that reply as speech, nor does it position the feature as a replacement for a qualified interpreter in medical, legal or other consequential settings.

The launch has four clear boundaries, and each should shape a buyer’s expectations before purchase or rollout:

• Pixel 11 first, not every Android phone;
• ASL-to-English first, not all 50 training languages;
• cloud translation, so connectivity and service availability matter;
• consumer UI only, with no public developer API or open weights.

Teams building Android accessibility features today still need platform APIs, user research and their own supported input paths. The Android app-generation guide can accelerate implementation, but it cannot substitute for Deaf-led design and evaluation.

Built with the Deaf community, not merely tested on it

Google says Deaf contributors shaped the project from initial concept through collection, evaluation and impact review. The work was initiated by a Deaf Googler, and the company formed an AI Sign Language Advisory Committee with global Deaf organizations and subject-matter experts.

That process is not decorative. A model can optimize sentence similarity while missing community priorities such as whether an interface treats signing as a first-class language, how consent works during data collection, what failure feedback feels respectful, and which language or dialect should be supported next.

The standard to watch is continuing power, not launch-day consultation: whether the committee can influence rollout, documented limitations, dataset governance and decisions to pause deployment when a subgroup is underserved.

Bottom line

SL2T’s most important achievement is product placement. Google did not ship a separate “AI sign language demo”; it put ASL input into the keyboard and conversation tool Pixel users already use. The landmark pipeline also makes a meaningful privacy trade: discard raw video locally while accepting that geometric movement data still goes to a server.

The breakthrough remains bounded. Pixel 11 supports one language direction, errors can change names, detail and tense, and there is no API or fully on-device translator. If Google expands the device and language coverage while preserving Deaf-led governance, SL2T could become the accessibility equivalent of voice dictation rather than another impressive model that never leaves the lab.

Related on explainx.ai

• Google DeepMind Magic Pointer: body movement as cursor control
• Augmental MouthPad: tongue-controlled computer input
• Neuralink’s telepathic wheelchair control demo
• Coral edge AI and on-device inference
• How to read AI benchmark claims
• Google AI Studio Android app generation
• GPT Realtime voice models and accessible conversation interfaces

Official source: Google DeepMind — Putting sign language AI into users’ hands

Device, language, benchmark and privacy details reflect Google’s August 12, 2026 launch materials. Pixel feature availability can vary by region and software rollout; verify the current device support and privacy notice before relying on the feature.