Voice-first AI presence

A text interface lets both parties pause without explanation. Speech is less forgiving. Every delay sounds like hesitation or absence. Talking over someone feels different from producing a long paragraph. A response that cannot be interrupted becomes a small captivity. When an agent speaks, timing becomes visible behavior.

Modern systems expose capabilities needed for a more conversational loop. OpenAI describes its Realtime API as direct speech-to-speech processing designed for low-latency interaction. Google’s Live API documentation exposes voice activity detection, interruption handling, turn-coverage settings, and non-blocking tool behavior. These are platform capabilities, not proof that a particular experience feels natural. They reveal the design surface.

Our thesis is that voice-first AI should be judged as a participant in turn-taking, not as a text model with audio output. The voice itself matters, but the social mechanics around it matter more.

Interruption is especially revealing. Google’s documentation notes that voice activity detection can cancel ongoing generation and that clients should discard buffered audio after interruption. This implementation detail becomes an experiential rule: the system must stop sounding as soon as it stops holding the floor. A delayed client can make an interruptible model feel stubborn.

A voice agent in a phone call is invited to treat nearly every utterance as addressed to it. An agent in a room, game, vehicle, or shared workspace faces a harder question: was that speech for the agent, for another person, or simply thinking aloud? Always answering turns presence into intrusion.

Proactivity features point toward this problem. Google documents a mode in which a live model can decide not to respond when input is irrelevant. But relevance is not only semantic. It can depend on wake cues, gaze, distance, role, the current activity, and the cost of interruption. This is where voice and spatial agency meet: who is near, what they face, and what place they occupy can help define whether the agent has the floor.

Silence should not be mysterious. The interface can signal “listening,” “not addressed,” “working,” or “uncertain” without speaking a sentence. Presence is not maximum output. It is the right level of availability for the moment.