Where the Second Goes

People on the phone expect an answer after a pause of about a second. Much longer and they say "hello?", start repeating themselves or talk over the reply, and the call stops sounding like a conversation.

So the useful question about a voice agent is not how clever its model is. It is how that second is spent.

The three parts of a pause

On 9 October 2026 we took a real call to our demo, running on cloud models, and split the gap between the caller's last word and the agent's first sound into its parts.

One call, one moment, cloud models. It is a breakdown, not a benchmark, and the delivery page shows the same figure with the same caveat.

Knowing the caller has finished

The first 0.48 seconds are spent waiting for silence to mean something.

A person who pauses in the middle of an address has not finished, and a person who says "yes" and stops has. The agent has to tell those apart from the sound alone, without the face and the breath that a human operator reads. Make the wait too short and the agent interrupts people mid-thought. Make it too long and every answer starts late. The same machinery decides when the agent should stop because the caller cut in, which is the subject of the caller who interrupted, and when a "mm-hm" is just a sign the caller is listening.

This is configuration, not hardware. It moves with your callers and your scenarios, and it has to be tuned on real calls rather than on a quiet office line.

The model's first sentence

The next 0.54 seconds belong to the language model, and the important word is "first". The agent does not wait for the whole reply. It needs the first sentence, hands it to the voice, and keeps writing the rest while the caller is already hearing the beginning.

That is why the model's speed to the first sentence matters far more than its speed on a long answer, and why a short first sentence ("Got it, a two-room flat.") is a design choice and not a mannerism. It is also the part that changes most when you move off cloud models: a local model on your cards can be faster or slower than this, and it slows down when several callers stop talking at once. We wrote about what that means for the server in what a voice agent runs on.

The first sound

The last 0.07 seconds are the synthesiser starting to speak. It is the smallest part on our call because the voice begins playing as soon as the first few words exist rather than after the full sentence is rendered.

A local synthesiser can lose that advantage if it has to finish a phrase before it can play it. That is a property of the model you choose, and it is worth checking before you choose it.

What happens to the numbers on your own servers

When the models leave the cloud, every row of the table becomes a question again.

The end-of-turn part stays roughly where it is, because it is mostly configuration, but the network part changes with where the server sits and how the call reaches it. The model's first sentence is set by your model on your cards and by how many callers are waiting at the same moment. The first sound is set by the synthesiser you pick. And a phone call adds its own delay on the way, which is part of what we measured in two mobile legs.

So we do not promise the demo's 1.09 seconds for a fully local install. We promise to measure it: on rented hardware first, with your scenarios and your real peak, before the server is bought.

How to write it into acceptance

A single good call proves nothing about the hundredth call of a busy morning. The criterion we propose has three parts: a median and a 90th percentile, from the end of the caller's phrase to the agent's voice, under an agreed number of simultaneous calls, on your real line. A typical line in the spec reads "median answer no slower than 1.5 seconds".

Written that way, "the agent is slow" stops being an impression and becomes a number both sides measured the same way.