How long do people wait for each other?
Hardly at all. A 2009 study led by Tanya Stivers at the Max Planck Institute for Psycholinguistics, published in PNAS, timed turn-taking in ten languages, from Danish and Japanese to the Mayan language Tzeltal, and found the typical gap between one person stopping and the next starting was around 200 milliseconds in every one. People start planning their reply before the other person has finished.
Phone networks are built around the same sensitivity. The ITU's recommendation G.114 advises keeping one-way delay on a voice line under 150 milliseconds for most conversations. A voice AI has to listen, think and speak inside a window that a phone line is only asked to carry sound across.
Where do the milliseconds go for a caller in Perth? (example figures)
A bike shop owner in Perth is on a voice call with an AI whose servers are in the United States. These are example figures for one reply.
The biggest share here is not the internet. It is the half second the system waits to be sure the owner has stopped, because a shorter wait means cutting in on people who pause to think. The model's time to its first sound comes next. Moving the servers to Sydney would trim about 160 milliseconds from this example, from 1,120 to roughly 960, which helps without changing the shape.
Why does distance add delay?
Nothing travels faster than light, and light in optical fibre is slower than in open space, at roughly 200,000 kilometres a second. Perth to Los Angeles is about 15,000 kilometres in a straight line, so a round trip cannot take less than 150 milliseconds before any computer has done any work, and undersea cables do not run in straight lines. Perth to Sydney is about 3,300 kilometres, a minimum of 33 milliseconds there and back. For Australians, where the servers sit is part of how a voice service feels.
What else makes a voice reply feel slow?
Unevenness, mostly. A reply that always arrives at the same pace is easier to talk with than one that is sometimes instant and sometimes takes two seconds. Talking over each other makes it worse: if the AI keeps going after you start speaking, the conversation stalls while you both stop. And a long answer feels slow even when it starts quickly, because there is no polite place to cut in.
Why does a call with Cora need to be quick?
Because CoraVision is a spoken conversation, latency is part of what you hear: you talk to Cora out loud, and each reply has to arrive at the pace of talk. CoraVision is a desktop app for Mac and Windows, and the onboarding call runs about 25 minutes. As with any voice call, your own connection plays a part, so a steady one, without a big download running alongside, helps the conversation flow.
Questions
What is a good latency for a voice AI?
The human benchmark is the 200 milliseconds that Stivers and colleagues measured across ten languages in 2009. Most voice AI cannot reach it yet, because it must first be sure you have stopped, and the nearer it gets, the more natural the call feels.
Is latency the same as internet speed?
No. Speed, or bandwidth, is how much data arrives each second, and latency is how long one piece takes to make the trip. A fast NBN plan can still have high latency to a server on the other side of the world, and a voice call cares about latency far more than bandwidth.