The total and the wait are different sums.

A conversational turn is a chain of stages, and there are two ways to add it up. The total is every stage's whole duration — the number on the dashboard. The wait is what the caller experiences: a stage that streams contributes only its time to first output, because everything after that happens while they are already listening; a stage that does not contributes all of it, because nothing downstream can begin.

The engine is free and needs no account. So a pipeline that spends 2.07 s can have the caller waiting 660 ms. The model runs for 900 ms and contributes 220; end-of-turn detection runs for 250 and contributes all 250. The silence detector is the bigger share of the wait — 37.9% against the model's 33.3% — at a quarter of the model's duration. And one blocking stage collapses the streaming around it: turn streaming off on the model alone and the wait goes from 660 ms to 1.34 s while the total does not move at all. Buffers count twice, because audio crosses them going in and coming out. Every lane ships a worked example that costs nothing. Writing the report needs credits.