What Sub-Second Response Actually Means on a Real Call
Latency in voice AI is usually quoted the way engineers like to quote it: model time, token time, a number from a benchmark. A caller cannot hear any of that. What a caller hears is the silence between the moment they stop talking and the moment something comes back. That gap is time-to-first-response, and it is the only latency number that changes how a conversation feels.
On a phone call, silence is the only feedback channel there is. No typing indicator, no spinner, no face. Past about a second of nothing, people assume something broke. They say hello again. They repeat the question. Then the reply finally arrives on top of the repeat, both sides talk at once, and the turn collapses into apologies.
When the clock starts
The clock does not start when the caller finishes their sentence. It starts when the system decides they have finished, which is a judgement call about a pause. Cut in too early and you interrupt someone who was thinking. Wait too long to be safe and you have spent your whole budget before a single word is generated.
Everything after that has to fit in what is left: recognising the speech, understanding it, looking up whatever the answer depends on, generating the reply, turning it into audio, and pushing that audio down the line. Which is why lookups have to be fast enough to be invisible. Memory, the shared customer record behind every Cirel call, answers in under 200ms precisely so that knowing who is calling never costs you the turn.
Why the PSTN part matters
Most voice demos run in a browser tab, and a browser tab is the easy case. The audio goes straight over the internet, in a modern codec, with no carriers in the middle. A real phone call is a different animal: it crosses the public telephone network, gets transcoded into narrowband, passes through jitter buffers and carrier hops, and picks up delay at every step.
That is why we measure sub-second time-to-first-response on real PSTN calls rather than in a browser. It is the harder number and the honest one. If a vendor quotes you latency, the question worth asking is simple: measured on what? A WebRTC demo can be twice as fast as the same stack on an actual telephone line.
There is one more thing the number does not capture. A fast reply that is filler still costs the caller a turn. Answering "let me check that for you" quickly is not the same as answering the question quickly. Sub-second matters because it buys back the room to say something real.
Easiest way to judge it is to ring one yourself. Book a demo and listen to where the silences land.