Google DeepMind Ships Sign Language AI to Pixel 11 Users

Google DeepMind shipped SL2T (sign-language-to-text) on August 12, 2026, embedded in Gboard and Live Transcribe on Pixel 11—the first sign language AI in a mainstream consumer product. The model translates American Sign Language to English in real time across any text field: web search, messages, Gemini, Live Transcribe. Pixel 11 ships August 20 at no added cost. The addressable market: 70 million Deaf and hard-of-hearing people globally, 200+ sign languages, none of whom had a voice-dictation equivalent until now.

The architecture uses a two-stage pipeline with a hard device/server boundary. Stage one runs entirely on-device: MediaPipe Holistic extracts 130 geometric landmark coordinates across face, hands, arms, torso in real time. Raw video is discarded on-device—it never leaves the phone. Stage two transmits only the lightweight coordinate sequence to Google's servers, where SL2T generates streaming English text. Sending coordinates instead of video cuts bandwidth and eliminates the core privacy objection to camera-based input.

SL2T drops the gloss layer that dominated prior sign-language pipelines. Gloss systems produced intermediate per-sign tokens before generating text; DeepMind argues glosses miss non-manual markers (facial expressions, mouth movements, spatial grammar) and cap vocabulary to labeled signs. SL2T maps landmarks directly to text—the move that shifted spoken-language MT from phrase tables to neural sequence-to-sequence. Translation quality now scales with data, not fixed annotation vocabulary.

Training corpus: 100,000+ hours across 50+ sign languages, roughly 25% in ASL. Multilingual joint training was deliberate—the model learns structural priors across languages. DeepMind reports it outperforms single-language baselines in internal testing. On FLEURS-ASL (sd-test), SL2T scores 70 BLEURT zero-shot, significantly higher than any published result. Zero-shot matters: the model generalizes without task-specific fine-tuning.

Production hardening went beyond benchmark optimization. DeepMind addressed: streaming latency (tuned but not disclosed); hallucination when the camera sees non-signing movement; left-handed signer recognition (10% of signers); one-handed accuracy, unavoidable on a phone where the other hand holds the device. None appear in FLEURS-ASL. They are the gap between benchmark and shipped.

At launch, the model handles ASL to English only. It struggles with rare signs and rapid fingerspelling. Google has not released error rates by signer demographic beyond left-handed and one-handed cases. DeepMind assembled an AI Sign Language Advisory Committee (AISLAC) from Deaf organizations and published an impact report listing prohibited contexts: healthcare, legal/law enforcement, academic evaluation, employment interviews, government hearings. The model is not a substitute for certified human interpreters.

SL2T is a functioning split-inference pipeline for body-pose input. The decision to discard raw video at the device boundary trades original footage for a lighter compliance surface—the design choice with the most consequence. Watch whether on-device landmark extraction becomes the latency floor as Google expands to lower-tier Android hardware beyond Pixel 11.