Putting sign language AI into users’ hands

Source: DeepMind Date: 2026-08-12 URL: https://deepmind.google/blog/putting-sign-language-ai-into-users-hands/

Summary

DeepMind’s SL2T is a sign-language-to-text model trained on over 100,000 hours across 50+ sign languages, using on-device pose tracking (MediaPipe Holistic) to translate video directly to text without an intermediary representation. It hits a zero-shot 70 BLEURT on the FLEURS-ASL benchmark and now powers ASL-to-English dictation in Gboard and Live Transcribe on Pixel 11.

Implications

Feeds local-model-feasibility as a case study: a specialized multimodal model (pose input, not raw video-to-LLM) doing real-time, on-device translation with hallucination suppression built in — the pattern of narrow, efficient models shipping alongside general frontier ones rather than being subsumed by them.