A caller waits before the AI's greeting, but the delay began before the application had a session. Optimising speech generation will not remove time spent establishing the telephone connection.
For engineers measuring voice latency, separate call setup from response generation. The caller experiences both, but the responsible components and useful repairs differ.
What is SIP setup delay versus AI response delay?
Setup delay covers the relevant telephone routing and connection stages before the conversation becomes available. AI response delay covers the later processing required to detect input, generate a response and deliver audio.
Define the boundaries explicitly for your measurement. Ringing, early media, agent startup and the first greeting can overlap differently across providers and deployment models.
Capture a first-contact timeline
Use an authorised test call and collect events from the actual route. The following worksheet avoids assuming that one application timer represents everything the caller waited through.
| Event boundary | Potential work inside it |
|---|---|
| Dial to provider receipt | Upstream carrier route |
| Provider receipt to application connection | SIP routing, dispatch and setup |
| Application connection to ready session | Worker and model preparation |
| Ready session to generated greeting | Application and speech processing |
| Generated audio to caller hearing it | Publication and media delivery |
The LiveKit SIP reference helps distinguish signalling evidence from later media evidence. Verify which events are available in the actual stack before naming a metric.
Compare the same entry route
A direct call to a provider test number may skip the business PBX, menu and overflow timeout. If the public route is slow, measure that complete route rather than compare it with a shorter path and blame the AI.
Test cold and already-running application conditions separately where relevant. A first call after deployment can have a different preparation path from later calls, so keep the configuration and workload visible.
The overflow timing guide covers waiting before the AI receives the call. The browser-versus-phone audio guide explains another source of misleading comparisons.
Repair the boundary supported by evidence
If routing consumes most of the wait, inspect the responsible provider or PBX rule. If the session is ready but audio generation is late, inspect the application pipeline. If output exists but the caller hears it later, inspect media delivery.
My voice platform contribution provides engineering context, not a universal latency benchmark. Report measured intervals with sample counts and known gaps rather than claiming a single response-time number describes every stage.
Can an AI model change fix a long ring interval?
It cannot remove waiting that occurs before the relevant application stage begins.
Should the first greeting and later turns use the same metric?
Define them separately because setup and preparation can affect the greeting differently.
What should we measure first?
Capture the public-route timeline and configuration before requesting a voice AI performance review.