Google DeepMind launches sign-language-to-text model on Pixel 11

Google DeepMind has released SL2T, a model that translates American Sign Language directly into English text, and baked it into Gboard and Live Transcribe on Pixel 11 devices. For the first time, deaf and hard-of-hearing users can sign to their phone instead of typing or dictating. The model was built with the deaf community — guided by deaf Google employees and a Sign Language AI Advisory Committee — and the company says it plans to extend support to 50 additional sign languages.

Under the hood, SL2T takes a different approach from earlier attempts that tried to break signing into discrete components. Instead, it captures the simultaneous motion of hands, body, and face as a continuous stream of body poses and processes it end to end. Sign languages are full natural languages with spatial grammar and facial expressions that carry syntactic weight, not just vocabularies of gestures. The researchers say the field's main technical blocker for years was the mistaken assumption that sign language is simply "words on hands."

Privacy is handled through a hybrid architecture. Pose tracking runs on-device; only the pose representation leaves the phone for cloud-based translation. Raw video never exits the device — a critical consideration for personal communication. The trade-off is practical: the translation step demands more compute than a phone can supply. Google has not published official latency figures, but claims the model is tuned for real-world use, including one-handed signing while the other hand holds the phone.

On benchmarks, SL2T hits state-of-the-art scores on standard academic tests. The more telling signal is its optimization for non-ideal conditions: shifting camera angles, poor lighting, fast signing, and one-handed signing. Early users report that existing captioning tools break down when conversation speed picks up; if SL2T holds up in live dialogue, that is the real breakthrough. No full technical report or code has been released yet, and the model remains Pixel 11-only for now.

The "with the community, not for the community" framing appears in every official statement. Deaf Google employees and the advisory committee were involved from problem definition through data collection to usability testing. This is not just ethics: without internal understanding of spatial grammar, cultural norms, and daily needs — such as one-handed signing — a model like this is set up for practical failure even if benchmarks look good. The planned expansion to 50 sign languages will require repeating that process for each language and deaf culture separately.