Agent Readout

Voice AI workload profile

Use these stats if you build synchronous voice loops.

Median latency
520-800ms
Streaming
Bidirectional via WebRTC beta

Guidance

  • Use short contexts (under 2k tokens) for better conversational pacing.
  • Silence detection threshold 120ms; adjust for non-English flows.
  • Fallback to text-only endpoint if voice infra unavailable (rare maintenance windows).
ModeHumanAgent