Turn-Taking Prediction

Turn-taking prediction studies when a speaker is likely to finish or continue, who may speak next, and whether an entry is a floor change, backchannel, or interruption. Systems use acoustic and prosodic cues such as pauses and intonation, often with voice activity projection, while other approaches add words, gaze, facial motion, speaker roles, or dialogue context. Current work combines streaming speech recognition with turn-state prediction and trains models using supervised, self-supervised, or synthetic conversational data. A central challenge is acting at the right time across noisy, overlapping, and multi-party conversations: early responses risk interruptions, while cautious systems may miss opportunities, and benchmarks do not always capture human judgments of appropriate timing.