Sign-Language-to-Text (SL2T)

Sign-Language-to-Text (SL2T) is a machine-translation model developed by the Google DeepMind Sign Language Team. Its first version, SL2T 1.0, was unveiled on 12 August 2026 and translates American Sign Language (ASL) into English text.

The model is a transformer-based network that conditions on a sequence representing a signer's movements and generates a text translation autoregressively. Its input is not a raw camera feed but a sequence of 2D landmark coordinates for 130 key points of the face, hands and body, extracted frame-by-frame using MediaPipe Holistic. It was trained on more than 100,000 hours of data across more than 50 sign languages, with roughly a quarter of the data in ASL.

Deployment and privacy

Video is processed on the user's device and then immediately discarded — only the abstract landmark representation is sent to Google's servers. The model runs in a hybrid mode: on-device pose detection and server-side translation. SL2T 1.0 is initially available on Pixel 11 phones in Gboard (text dictation) and Live Transcribe (conversation responses) and is currently free.

Governance (AISLAC)

Deployment is guided by the AI Sign Language Advisory Committee (AISLAC), an advisory body convened by Google that brings together Deaf advocacy organizations, academic and cultural institutions and interpreters (including the National Association of the Deaf, RIT/NTID, DPAN and the World Federation of the Deaf). The Google DeepMind team and AISLAC co-authored the "Joint Impact Report for SL2T 1.0", detailing the model's capabilities and limitations. High-stakes uses (medical, legal, academic, and multi-party conversations) are explicitly out of scope for this release.