Building sign language conversations from individual signs

Sign language users often rely on spoken or written language intermediaries in conversational AI, limiting accessibility. This work addresses the scarcity of sentence-level sign video data by constructing continuous sign conversations from large collections of isolated signs labeled with motion primitives. SignaVox-W provides the largest isolated-sign vocabulary to date; SignaVox-U is a 3D conversation dataset built from it. The system uses a retrieval-guided translator to bridge spoken and signed language structures, and BRAID—a diffusion Transformer—aligns durations and fills co-articulatory gaps between independently collected clips. The resulting SignaVox model generates full-body, hand, and facial responses directly from prior signing context, without relying on text or glosses at inference time.