Audio streams in while audio streams out. Between them, the system detects turns, updates the conversation, calls tools when needed, and reacts when a person interrupts.
Two musicians listen and respond continuously. Timing, pauses, and interruptions are part of the performance.
A persistent session carries small audio and control events in both directions instead of waiting for one complete recording.
Realtime is a session, not one request. State evolves as audio and control events arrive.
Designed for realtime media transport and resilient playback.
Direct event stream when your backend controls the audio pipeline.
Connect voice agents to conventional calling infrastructure.
Network jitter creates gaps. Aggressive turn detection cuts people off. Slow tools stretch silence. Echo can look like a new speaker. Playback must be truncated when the user interrupts so session state matches what they actually heard.
Capture audio continuously.
Detect turns and produce responses.
Handle tools, errors, and interruption.