The call connects, but the first seconds are silent. That silence may come from media setup, agent preparation, an intentional greeting rule or output that was generated but never played to the caller.
For engineers investigating first-call silence, collect a timeline before adding a filler greeting. A workaround that speaks earlier can conceal a broken input or output path without making the session ready for conversation.
What can cause silence at call start?
Connection, media availability, agent readiness and audio playback are separate events. The application may report a ready session before the caller can hear the published output.
Define whether the assistant is supposed to greet immediately or wait for the caller. An intentional wait is a product policy; an unintended gap needs evidence about the responsible boundary.
Trace the first audible response
Use a controlled telephone call and compare what the caller heard with application and provider events. Preserve the actual configuration used by the test.
| Observation | Boundary to investigate |
|---|---|
| No session is dispatched | Routing and agent selection |
| Session starts late | Worker preparation and dependencies |
| Output generation starts late | Application greeting logic and model readiness |
| Audio is published but not heard | Media delivery and negotiation |
| Caller speech is also missing | Incoming audio and subscription path |
The LiveKit SIP troubleshooting reference separates signalling from media investigation. Use it where that platform is part of the route and inspect other providers at their own boundaries.
Compare cold and repeated calls explicitly
If the first call after deployment is slower, record preparation events and resource conditions. Do not assume every later call will share the same readiness path, especially after scaling or restart.
A fixed wait inserted before greeting may make one test look orderly while adding unnecessary delay elsewhere. Prefer a documented readiness condition and bounded failure response where the application supports it.
The setup-versus-response timeline helps identify which interval actually contains the silence.
Verify that the caller can respond immediately
After the repair, test a caller speaking over or just after the greeting. Confirm that incoming audio is available and the first useful request is not lost.
My voice platform contribution provides relevant context, not a promised startup time. Retest the actual telephone route and record the tested release, provider arrangement and known limitations.
The phone audio regression checklist can keep this scenario in later change reviews. A first greeting that sounds correct is one acceptance condition, not proof that transfers or business writes work.
Should we solve silence by adding a longer greeting?
First identify the missing readiness or media boundary. Extra speech can hide the symptom without fixing the cause.
Is first-call silence always a cold-start problem?
No. Greeting policy, dispatch and media delivery can create similar symptoms.
What evidence should we collect?
Capture connection, readiness, output and caller-audible events for the voice AI technical discussion.