Google DeepMind launches SL2T sign language recognition model on Pixel 11 devices

Google DeepMind has launched SL2T, a groundbreaking sign-language-to-text model debuting on the Pixel 11 smartphone. Operating within Gboard and Live Transcribe, SL2T allows Deaf and hard-of-hearing users to sign directly into their device's camera instead of typing. The model translates body movements directly to English text, bypassing restrictive intermediate "glosses." To protect privacy, on-device MediaPipe Holistic software converts video into abstract geometric coordinates, discarding raw footage immediately. However, Google and its Deaf advisory committee explicitly prohibit SL2T's use as a substitute for certified human interpreters in high-stakes legal, medical, or educational settings.

SL2T sign language model launch

• ▪Google DeepMind launched its sign-language-to-text model, SL2T, inside Gboard and Live Transcribe as part of the Pixel 11 smartphone announcement on August 12, 2026.
• ▪The Pixel 11 smartphone is scheduled to ship on August 20, 2026, making it the first device where users can sign directly into text fields.

Gloss-free translation approach

• ▪Google DeepMind's SL2T model translates body movements directly to English text, bypassing the intermediate word-for-word English labels known as glosses.
• ▪By avoiding glosses, the SL2T model is able to capture non-manual markers, spatial grammar, and classifier handshapes that are invisible to gloss-based systems.
• ▪The SL2T model achieved a zero-shot score of 70 BLEURT on the FLEURS-ASL benchmark and a score of 85 BLEURT on the PRESTO-ASL benchmark.

On-device privacy architecture

• ▪An on-device model built on MediaPipe Holistic converts the camera feed into 130 geometric landmark coordinates and discards the raw video immediately.
• ▪Researchers have argued that dense facial landmark coordinates could theoretically constitute biometric identifiers under biometric data privacy laws.
• ▪A transformer neural network on Google's servers processes the coordinate sequences to generate English text autoregressively without retaining user input logs.

ASL accessibility impact

• ▪The SL2T model allows Deaf and hard-of-hearing users to sign directly into their smartphone's camera to perform web searches, draft messages, and query Gemini.
• ▪Google DeepMind and the AI Sign Language Advisory Committee explicitly state that SL2T 1.0 must not be used as a substitute for certified human interpreters in high-stakes settings.
• ▪The joint impact report states that SL2T 1.0 does not satisfy legal obligations for reasonable accommodations under the Americans with Disabilities Act (ADA).

Sign language AI development challenges

• ▪Google DeepMind trained the SL2T model on over 100,000 hours of data spanning more than 50 sign languages to learn shared structural patterns.
• ▪Sign language AI development has lagged because sign languages convey meaning through simultaneous movements of the hands, head, face, arms, and torso.
• ▪Prior sign-language gloves failed to scale because they read hand shapes alone and ignored the rest of the body.