Google DeepMind has released SL2T, a sign-language-to-text AI model that lets people sign directly to their phone instead of typing, marking what the company describes as the first deployment of sign-language AI in mainstream consumer software.
Where it ships
The model launches inside two existing Google apps: Gboard, where users can sign to search the web, draft messages, or ask Gemini a question, and Live Transcribe, where Deaf and hard-of-hearing users can sign their side of a spoken conversation instead of typing replies. Both features are free and arrive first on the Pixel 11, which ships August 20, with DeepMind saying more devices will follow.
At launch, SL2T translates American Sign Language (ASL) into English. DeepMind said additional sign languages are planned but did not give a timeline.
How it was built
According to DeepMind, SL2T was trained on more than 100,000 hours of data spanning over 50 sign languages, with roughly a quarter of that data in ASL, drawing on varied signers, dialects and skill levels. On the FLEURS-ASL benchmark, the model scored 70 BLEURT in zero-shot testing, which DeepMind said is well above any previously reported result for the task.
The project originated with Sam Sepah, a Deaf Googler, and was shaped throughout by an AI Sign Language Advisory Committee that includes Deaf-led organizations and linguists, according to the company. DeepMind also disclosed remaining weaknesses: the model can still misread rare signs, fast fingerspelling, passive grammatical constructions and tense, and it is not yet fully optimized for left-handed or one-handed signing done while holding a phone.
Why it matters
Sign-language recognition has long lagged behind speech and text AI because usable training data is scarce and sign languages vary widely by region. Putting a model directly into a keyboard and a transcription app, rather than a standalone research demo, gives DeepMind a real-world test of whether the accessibility technology holds up outside the lab, and could push rival phone makers to add similar tools.
The release follows years of smaller sign-language AI prototypes; DeepMind’s own research into gloss-free translation, which skips the intermediate step of tagging individual signs, underpins the jump in benchmark accuracy.