SignGemma

SignGemma is an open model in Google's Gemma family that translates sign language into spoken text in real time. Unveiled at Google I/O 2025, it pairs a vision transformer that tracks hands, motion, and facial expressions with a compact language model that writes the English translation. It was trained on more than 10,000 hours of annotated sign language video matched with transcripts.

The model is built to run fully on-device on phones, tablets, and laptops with no internet connection, which matters for privacy since signers' videos never leave the device. It performs best on American Sign Language to English today, with other sign languages planned. Google opened an early-testing interest program for developers building accessibility applications.