Active speaker detection (ASD) aims to identify which person in a scene is currently speaking. Recent models improve accuracy by jointly processing all visible faces, but this typically happens offline, which rules out real-time use.
CAST-ASD is presented as a causal joint model that applies attention over both speakers and time. Because it only relies on past and current information, it can make predictions online while still benefiting from joint face processing.
The source is an arXiv announcement whose abstract is truncated, so no quantitative results are reported here. The contribution described is architectural: bringing causal, online operation to joint active speaker detection.