Key takeaways
- Measure from the click, not from when audio connects.
- Report the median and a high percentile, not just the best case.
- Warm pools of ready agents and browsers shorten it.
What adds to time to first word
- 1The call page loads and asks for the microphone.
- 2The browser connects to the real-time audio room.
- 3An agent process joins and, in live product mode, a cloud browser starts.
- 4The agent's greeting is generated and the first audio plays.
Why the first seconds matter
People decide quickly whether something works. Silence after a click reads as a failure, and some will close the tab. Starting the greeting before every system is fully ready, and filling short gaps naturally, keeps them on the call.
TalkDemo's target
The PRD sets targets of a median time to first word under three seconds and under five seconds at the 95th percentile. These are design targets for a product in private pilots, not measured results. The agent and browser start in parallel to help meet them.
Time to first word vs reply latency
- Time to first word happens once, at the start of a call, and includes setup.
- Reply latency happens after every question and covers listening, thinking and speaking.
- Both shape how human the call feels, and both are worth reporting as percentiles rather than averages.