Google DeepMind SL2T moves sign-language AI toward a split mobile runtime

Google DeepMind introduced SL2T, a sign-language-to-text model starting with ASL-to-English on Pixel, with related social posts explaining benchmarks, privacy, and Deaf-community input.

Retrieval answer

Google DeepMind announced SL2T, a sign-language-to-text model reaching Android/Pixel product surfaces, starting with ASL-to-English on Pixel 11. The cluster includes a primary blog and official social thread with model claims, benchmark/privacy details, and Deaf-community input. This is a material accessibility AI release with concrete product consequences and no prior coverage match.

A privacy-aware flow from camera input through on-device pose landmarks to server translation and streaming text in Gboard and Live Transcribe. — New Runtime synthesis: SL2T separates on-device landmark extraction from server-side sign-language translation.New Runtime synthesisOriginal source ->

Field note

Google DeepMind introduced SL2T, a sign-language-to-text model starting with ASL-to-English on Pixel, with related social posts explaining benchmarks, privacy, and Deaf-community input.

Why it matters

New Runtime view

SL2T is a split accessibility runtime: local feature extraction narrows the media boundary before a frontier model performs language translation. The consequence is not only better benchmark translation, but a product architecture for camera-based assistants where raw human motion should not automatically become cloud media.

Mechanism: On-device MediaPipe Holistic extracts body, face, and hand landmarks; Gemini 2.5 Pro translates the landmark stream into text.

Architectural boundary: Primary perception and privacy-sensitive video handling sit on Android/Pixel while semantic translation is server-side.

Measured consequence: DeepMind reports BLEU improvement from 15.0 to 21.4 versus the prior best on YouTube-ASL/How2Sign-style datasets.

What remains open

Research preview, not a certified interpreter replacement.
Signer diversity, dialect coverage, latency, consent UX, and accessibility liability remain open.
Landmarks reduce raw-video exposure but can still encode sensitive expression patterns.

Sources

<https://deepmind.google/blog/putting-sign-language-ai-into-users-hands>
<https://x.com/GoogleDeepMind/status/1955645677943959647>

Evidence / sources

Recommendation

Google DeepMind introduced SL2T, a sign-language-to-text model starting with ASL-to-English on Pixel, with related social posts explaining benchmarks, privacy, and Deaf-community input.