Google DeepMind Launches SL2T Sign Language Dictation on Pixel 11
Google DeepMind’s SL2T brings ASL-to-English dictation to Gboard and Live Transcribe on Pixel 11, using pose landmarks to protect video privacy by design.
Summary
On August 12, 2026, Google DeepMind and Android introduced SL2T, a multilingual sign-language-to-text model powering what Google calls the first consumer deployment of sign language AI. Available at no additional cost in Gboard and Live Transcribe on Pixel 11, it initially translates American Sign Language into English, letting users sign searches, messages, documents, Gemini requests and conversational replies. More devices and languages are planned, addressing an estimated 70 million Deaf and hard of hearing people across more than 200 sign languages.
SL2T was trained on over 100,000 hours spanning more than 50 sign languages, roughly one quarter in ASL; joint multilingual training outperformed single-language models in Google’s experiments. On-device MediaPipe Holistic converts movement into pose landmarks, sends only geometric coordinates to servers and immediately discards raw video. Direct landmark-to-text translation avoids gloss annotations and their vocabulary and structural limits. SL2T scored 70 BLEURT zero-shot on FLEURS-ASL sd-test, above previously reported results, while work also targeted streaming latency, non-signing hallucinations, one-handed use and fairness for the 10% of signers who are left-handed. Errors remain with rare signs, rapid fingerspelling, passive constructions, classifier depictions and context-dependent tense. Deaf Googler Sam Sepah helped conceptualize the project, while Deaf partners, user studies, experts and the AI Sign Language Advisory Committee shaped data, evaluation, governance and a joint SL2T 1.0 impact report. Google plans additional sign languages, sign generation and frontier AI capabilities.
Positives
- Over 100,000 training hours across more than 50 sign languages helped multilingual SL2T outperform Google’s single-language models.
- 70 BLEURT zero-shot on FLEURS-ASL sd-test exceeds every previously reported score cited by Google DeepMind.
- Gboard and Live Transcribe provide free ASL-to-English signing input on Pixel 11 for searches, messages, documents, Gemini tasks and conversations.
- MediaPipe Holistic discards raw video immediately and sends only pose landmark coordinates to the translation server.
- The 10% of signers who are left-handed, plus people signing one-handed while holding phones, received explicit performance and fairness attention.
- AISLAC, Sam Sepah, Deaf organizations, partners, testers and experts directly influenced SL2T’s design, evaluation, deployment and impact reporting.
Risks & concerns
- Pixel 11 launches with only ASL-to-English support, while more devices and sign languages remain future commitments.
- Rare signs, passive constructions, classifier depictions and context-dependent tense can still produce omissions or mistranslations.
- Rapid fingerspelling confused “prey” with “grey” in a published FLEURS-ASL example.
- Pose coordinates still require server-side translation, although raw camera footage is discarded on the device.
- More than 200 sign languages serve an estimated 70 million Deaf and hard of hearing users, far beyond SL2T’s initial consumer coverage.