What is latency on an AI call?
- Latency
- Latency on an AI phone call is the delay between the caller finishing what they were saying and the assistant starting to reply, measured across recognition, reasoning and speech synthesis together.
What latency means in practice
Three stages each add their share. Recognizing the audio, working out the reply, and generating the voice.
Humans are sensitive to this in conversation. A pause that would be unremarkable in a text chat feels wrong on a phone call.
It is also uneven. A simple reply comes back quickly and one requiring a calendar lookup takes longer, and the variation is more noticeable than the average.
What people get wrong
Where 1.6 seconds goes
Here's an illustrative budget for a single reply. These are round numbers chosen to show the shape of the problem. They aren't measurements of GreetKeeper or of anyone else.
Your caller stops talking. The system waits about 500 milliseconds to be sure they're finished and haven't just paused for breath. Recognition takes 200. The language model needs 500 to start producing a reply. Speech synthesis takes 250 to generate the first audio, and the phone network adds 150 carrying it back. That's 1.6 seconds of silence from the caller's side.
For comparison, research on everyday conversation puts the typical gap between two human speakers at around a fifth of a second. Look at the budget again and you'll see the biggest single item isn't computing at all. It's the wait to decide the caller has finished, which is why latency and turn-taking can't be tuned separately.
Measuring it yourself with a voice memo
You don't need lab equipment. Put a demo call on speaker, record it with the voice memo app on a second phone, and have a normal conversation of ten or twelve turns. Include one question that needs a lookup, such as asking for the next open appointment.
Play the recording back and time each gap from the end of your sentence to the first sound of the reply. A stopwatch is accurate enough. Write all the gaps down, then look at the longest two. Averages flatter every system, and what a caller remembers is the one four-second silence in the middle of booking. Make the call from a cell phone with two bars if you can, since that's closer to your real callers than an office landline will ever be.
How GreetKeeper handles it
GreetKeeper publishes no latency figure, because we have not benchmarked one honestly and we don't print a number we have not measured.
Judge it on a demo with your own scenario. That is the only test that includes your network, your caller's line and the kind of question you actually get.
Where a pause is unavoidable, such as during a transfer, the assistant should say what is happening rather than going silent.
Latency questions
What delay do callers notice?
Sooner than you would expect, because phone conversation has tighter timing than any other kind. Rather than chasing a threshold, listen to a real call and judge whether it felt like waiting.
Does a longer answer take longer to start?
Not necessarily, since speech can begin before the whole reply is composed. What reliably adds delay is anything that has to be looked up first.
Is my internet connection part of it?
Your caller's connection matters more than yours, since they are the one on the line. A poor mobile signal adds delay no vendor can engineer away.
Related terms
Hear it take one of your calls
Two minutes, your own scenario, no card.