Full-Duplex Interaction
Bidirectional continuous communication where both parties can send and receive simultaneously.
Full-duplex interaction is a communication mode in which both participants can transmit and receive simultaneously — in contrast to half-duplex, where participants must alternate speaking and listening. The term comes from telecommunications, where a full-duplex phone call allows both parties to speak and hear each other at the same time, whereas a walkie-talkie requires push-to-talk and explicit turn-taking. In the context of AI interaction, full-duplex refers to a system where the model can perceive incoming audio, video, or text while simultaneously generating output, without being blocked by the input stream.
Full-duplex interaction is what distinguishes a natural conversation from a formal debate or a written exchange. In a phone conversation, interruptions, overlapping speech, and simultaneous back-channeling (uh-huh, mm-hm) are natural and informative — they signal attention, agreement, confusion, and intent in real time. A half-duplex AI system, by contrast, forces an artificial serialization of these signals: the user speaks, then the model speaks, then the user speaks again. This serialization discards the rich parallel communication channels that humans rely on in natural collaboration.
The concept matters in AI because most production real-time AI systems today are architecturally half-duplex, even when they feel responsive. A voice assistant that waits for the user to finish speaking before responding is operating in half-duplex mode: it has stopped processing input in order to generate output. Full-duplex systems like Moshi, PersonaPlex, and Nemotron VoiceChat process input continuously and can respond at any point — they do not wait for explicit turn boundaries. The key advance of interaction models is achieving full-duplex capability across audio, video, and text simultaneously, with the intelligence of a frontier model rather than the simplicity of a voice-optimized specialist.
Full-duplex interaction creates new challenges: the model must handle overlapping speech gracefully, decide when to interject versus yield, and manage the safety implications of a system that speaks proactively rather than only in response. These are qualitatively different engineering problems from half-duplex turn-taking, and they require architectural changes rather than merely faster inference.