The model maintains a continuous audio input stream and a continuous audio output stream. A decision layer picks one of speak, listen, pause, interrupt or call-tool many times per second. Deep-work queries (web search, multi-step reasoning) are delegated to a background reasoning model (e.g. GPT-5.5), whose answer is folded into the conversation without breaking its rhythm. The model tracks user pace, volume, tone and background noise so it can distinguish real turn-ends from short thinking pauses.
Addresses the stiffness of turn-based voice interaction: long pauses, unnatural interruptions, high latency of the STT → LLM → TTS pipeline, and the difficulty of holding a natural conversation with active listening.
