Google SignGemma: ASL to English text, fully on-device and open source

Google SignGemma is an AI model from Google DeepMind: it translates American Sign Language (ASL) into English text in real time, on phones, tablets and laptops. Everything runs locally, with no internet connection required. Unveiled at the I/O conference in May 2025 as part of the open Gemma family, it is aimed at deaf and hard-of-hearing people as much as at developers building accessibility apps. Access currently goes through a free early-testing program.

Hands, face, motion: the signals SignGemma reads

SignGemma pairs two components: a vision transformer that tracks hands, motion and facial expressions, and a compact language model that writes the English translation. Google trained it on more than 10,000 hours of annotated ASL video matched with English transcripts.

In practice, you sign in front of the camera and the text appears almost instantly: early internal tests mention latency under 200 milliseconds on some devices. That speed matters, because American Sign Language relies on the face as much as on the hands, which explains this full-body reading of the signer.

SignGemma, from the I/O stage to the testing program

Google introduced SignGemma at its I/O conference in May 2025, with a live demo by Gus Martins, Gemma product manager at DeepMind. Since then, access has gone through an interest form aimed at three groups: developers, accessibility researchers and members of the Deaf community.

The open release of the weights, originally expected by the end of that year, is still pending: on Google's developer forum, researchers were still asking for updates months later. The form remains the official way in.

A pocket translator built with privacy in mind

Local processing changes everything for something as personal as a signed conversation: nothing is sent to a server, and the video stays on the phone unless you choose to share it. That on-device design also helps in areas with poor connectivity, a rarity among AI translation tools, most of which depend on the cloud.

Human interpreters keep their place for complex situations: regional dialects, overlapping signs, cultural context. SignGemma targets everyday life instead, the classroom, the front desk, the video call (in other words, all those moments when no interpreter is around, which is most of them).

Frequently asked questions