Simon Michael Vuch Williams

TurnBench Exposes Turn-Taking Failures Across Conversation Types

A 30-hour triple-annotated benchmark separates end-of-turn and interruption decisions, showing that low-latency interruption detection still produces excessive false positives.