Turn-Taking in Dialogue

Turn-taking research examines how people and conversational systems manage the floor: detecting when a speaker is finished, recognizing backchannels, choosing who speaks next, and timing interruptions. Methods range from voice-activity and prosody models to streaming speech-language models that combine acoustic, semantic, and sometimes visual cues such as gaze; systems are trained with supervised data, synthetic dialogues, and reinforcement learning. Benchmarks increasingly test full-duplex and multi-party interaction, different conversational intents, and noisy or overlapping speech, while measuring timing alongside response content. A central challenge is making systems as context-sensitive as human listeners: they must avoid both premature interruptions and missed opportunities to speak, and current systems still struggle to match human timing reliably.