Evaluating Conversational Latency and Turn-Taking in Speech Models
October 12, 2026
An evaluation of six speech systems shows that GPT Realtime, Gemini 3.1 Flash Live, and Grok Voice respond at nearly every pause, but latency remains the primary bottleneck. Full-duplex models like Moshi are the only ones currently regulating speech timing to match human-like conversational turns.
HOW THIS AFFECTS YOU
●
builderYou should consider full-duplex architectures if natural turn-taking is a core requirement for your voice agent.
●
designerThe latency gap between silence-based and semantic endpointers directly dictates the UX of conversational pauses.