Google DeepMind has introduced SL2T, a sign-language-to-text model trained on more than 100,000 hours of data spanning over 50 sign languages, roughly a quarter of it American Sign Language. Rather than processing raw video, the system runs an on-device MediaPipe model to extract pose landmarks, which are the only data transmitted for translation, a design intended to protect user privacy. The model translates the landmark sequence directly into text without intermediate “gloss” annotations, an approach DeepMind says better captures sign language’s non-manual features and lets performance scale with additional training data.
