RSATalker: Realistic Socially-Aware Talking Head Generation for Multi-Turn Conversation

In plain English

A talking-head pipeline that drives a 3D facial mesh from speech, binds Gaussians to the mesh facets for rendering, and conditions the output on the social relationship between the speakers — blood versus non-blood, equal versus unequal standing — on the premise that the same sentence should be delivered differently to a manager than to a sibling. The authors also release a speech-mesh-image dataset annotated with those relationships, which is arguably the more reusable contribution. The abstract claims state-of-the-art realism and social awareness but reports no metrics, so the performance claim cannot be checked without the full paper.