DeepSign

Sign language is a natural visual language whose tokens are fine-grained body and hand motion — a language where video understanding and video generation must meet. DeepSign is a single foundation model that covers both directions of the task: video to text, translating live signing into spoken-language text; and text to embodied motion, generating natural, expressive signing from written or spoken language. It runs in real time on live streaming video, with latency low enough for natural two-way conversation, and achieves state-of-the-art performance in both directions of the task. Beyond accessibility, DeepSign is a proving ground for world models that both perceive and act. It has not been publicly released yet — more details coming soon.