Thanks — I hear you: you’re seeing a mismatch between the audio in Agent Insights and the audio from LiveKit Egress, and it sounds like the agent is missing parts of the conversation. Below I’ll explain the most common causes and give a short checklist of targeted diagnostic steps you can run. If you want, after those checks you can share a few details so I can help narrow it down. Why this can happen (short version) • Agent Insights shows audio as the agent actually “hears” after any agent-side processing (VAD, noise cancellation, BVC, end-of-turn detection). That processing can remove or suppress audio that still exists in the room-level egress recording. • Egress records room-level audio (mixed tracks or specific tracks depending on your egress config) and does not apply the agent pipeline filtering, so it can contain speech that the agent didn’t process. • Timing/end-of-call differences, aggressive VAD, or incorrect subscription to participant tracks can create the appearance that the agent “didn’t hear” something even though it’s in the egress file. Quick checklist to diagnose the mismatch 1. Compare timestamps ◦ Align a short segment from the egress file with the Agent Insights timeline to see exactly which timestamps differ. 1. Check the Agent Insights Trace view ◦ Look for spans labeled

vad

,

noise_cancellation

,

stt

, or

eou_detection

. These show where audio was filtered or where end-of-turn detection occurred. 1. Confirm egress configuration ◦ Are you using Room Composite (mixes all tracks) or Track/Participant egress (captures selected tracks)? Make sure egress includes the participant(s) you expect. 1. Verify which track the agent is subscribed to ◦ Ensure the agent is subscribed/pinned to the correct participant (for SIP calls, the SIP participant specifically) rather than “any” or a different participant. 1. Test with processing disabled ◦ Temporarily disable noise cancellation / BVC / aggressive VAD to see if the missing audio reappears in Agent Insights. 1. Inspect logs for warnings ◦ Look for messages like “VAD did not detect speech”, audio input failures, or high latency in STT/LLM processing that could cause missed turns. Targeted follow-ups I can help with • If you want, share: ◦ Whether the call is SIP or WebRTC, ◦ The agent’s VAD/noise-cancellation settings, ◦ A short time-aligned clip (egress vs agent insights) and the session or room ID. • I can walk you through reading the Trace view, advise how to pin the agent to a specific participant, or suggest the exact log entries to look for. If you prefer, tell me which diagnostic step you ran (or want to run) and I’ll guide you through it. Sources: Streamline troubleshooting with Agent Observability | Improved End-of-Turn Model Cuts Voice AI Interruptions 39% | Choosing the right noise cancellation for your agent | LiveKit | Universal Egress | Universal Egress