Add silence/turn-taking detection so Realtime Chat Agent knows when not to speak

Currently, the Realtime Chat Agent may try to respond to every detected speech, even in contexts where it should remain silent. For example: • When people in the environment are conversing with each other and the agent is not being addressed. • When the conversation clearly does not involve or require the agent. • When background chatter or noise is detected.

It would be useful if the agent could intelligently determine when it should speak vs. when it should stay silent, based on conversational turn-taking, speaker intent, or whether the input is directed at the agent.

Why this matters • Makes the agent feel more natural in multi-person environments. • Prevents unnecessary or awkward interruptions. • Aligns behavior closer to human conversational norms.

Possible directions • Add a configurable “silence policy” to decide when not to generate a response. • Use intent classification (e.g., addressed-to-agent vs. ambient speech). • Expose hooks so developers can override or implement their own silence logic.