Google DeepMind launches sign-language-to-text AI, powering ASL dictation on Pixel 11

Google DeepMind has introduced a new sign-language AI model that translates sign language directly into text, and shipped it as a real product feature on the Pixel 11 rather than a research demo.

What's new

The model, called sign-language-to-text or SL2T, is what DeepMind describes as "a breakthrough in quality and generality" that brings sign language AI to consumer products for the first time. Rather than processing raw video, SL2T "translates sign language inputs as points on the signer's body" — tracking body position data instead of full video frames, which DeepMind says enables the feature to run with on-device privacy protection.

The model now "powers sign-to-text dictation in Gboard and Live Transcribe on Pixel 11, starting with American Sign Language (ASL) to English," according to DeepMind's announcement. DeepMind says "more devices are coming soon, and additional languages will follow," indicating this is a first rollout rather than the full scope of the feature.

On the training and evaluation side, DeepMind says SL2T was "trained on over 100,000 hours of data across more than 50 sign languages — with roughly a quarter of the data in ASL," and that it "achieves a remarkable zero-shot score of 70 BLEURT, which is significantly higher than any previously reported score" on that benchmark. BLEURT is a machine-translation quality metric, and a zero-shot result — meaning performance on sign languages the model wasn't specifically fine-tuned for — of that magnitude would represent a meaningful jump over prior published sign-language translation systems.

Context

Google built SL2T in partnership with the AI Sign Language Advisory Committee (AISLAC), which DeepMind describes as bringing together "global Deaf organizations and subject-matter experts" under what the company calls a "participatory governance model." That framing signals an attempt to avoid a recurring criticism of past sign-language AI efforts: that they were built without meaningful input from Deaf communities and signers themselves, and often shipped as translation gimmicks rather than genuinely useful accessibility tools.

The announcement lands the same week Google shipped the Pixel 11 with an expanded set of Gemini-powered automation features, underscoring how much of Google's current hardware push is built around bundling frontier AI research directly into consumer devices rather than keeping it confined to research papers or developer APIs.

Why it matters

Sign language translation has been a long-standing, largely unsolved target for AI research — sign languages are visually and grammatically distinct from spoken/written languages, vary widely by region, and have historically lacked the scale of paired training data available for text-to-text translation. A body-tracking approach rather than raw video processing is notable both for privacy (it can run on-device without transmitting video) and for computational efficiency on a phone.

Shipping this directly into Gboard and Live Transcribe — tools ASL signers can use immediately for dictation and conversation transcription — moves sign-language AI from a research benchmark into daily assistive technology, competing directly with efforts elsewhere in the industry to build accessible AI interfaces. If the claimed BLEURT gains hold up under independent scrutiny, it would mark one of the more substantive accessibility-AI advances of the year, and a template other platform vendors are likely to be measured against.