Callers do not hear a median
The figure in the title is our median voice latency, measured mouth to ear. It describes the turn in the middle, and nobody on a call hears the turn in the middle.
A caller remembers the turn where the line went quiet long enough that they said “hello?” into it. That is the turn worth engineering for, and the median hides it.
Where the milliseconds go
Until the March release, each stage of a voice turn waited for the one before it. The rebuild runs recognition, reasoning and speech as one path, and most of the saving came from that overlap.
The largest share of a turn, about 170 milliseconds, goes to deciding that the caller has finished, and that is a guess. Guess late and every turn carries dead air. Guess early and the agent cuts in on a caller who has paused halfway through a booking reference.
The turn where a caller stops waiting
Between two people, the gap from one turn to the next is usually about a fifth of a second. Callers give an agent a little more room, and our median now sits near the edge of it, where a caller starts talking over the silence.
A call is not one turn. Say it has 14 agent turns, and one in ten runs past a second. The chance that the call holds at least one long silence is better than three in four.
Nearly all of our long turns wait on another system to find a booking, and no work on the speech path makes that system reply sooner. Meanwhile the caller asks whether anyone is there or starts the question again, and one slow turn has become two.
Speed is not what a caller notices. What they notice is the one turn where nothing happened and nobody told them why.
What the median actually buys
A person on a desk does not go quiet while a record loads. They say they are pulling up the booking, and the caller waits without minding, because the silence has been explained.
An agent can work the same way, and a fast median is what lets it. After a run of quick turns, a caller gives the slow one patience if told what is happening. That is what the number buys: room to spend where the work needs it.
On your own lines
Read the latency for each number on the voice dashboard, and look at the slow end, not the median.
Find the steps the slow turns sit on, and give each a short, true line to say before it goes quiet.
Search the transcripts for a caller saying “hello?” partway through a call, or asking the same question twice.
Test with references read the way callers read them, in groups and with pauses, and listen for the agent cutting in.
The four together will barely move the median. They move the turns past it, and those are the turns a caller describes to someone else after they hang up.
