Glossary · Models & serving
What is caller-perceived latency?
Caller-perceived latency
Caller-perceived latency is the silence a caller hears between finishing their sentence and hearing the agent reply. It's the delay that counts, because it's the only one the caller experiences.
On a call
The caller says "हाँ, कल दस बजे ठीक है" and stops. If the agent's "बढ़िया, मैंने book कर दिया" starts a moment later, the call flows. If it takes two seconds, the caller says "hello?" and both of them start talking at once.
AI agents from ₹1 a minute.¹
Explained
Here is how it works, in plain words.
Engineers measure lots of latencies, from the model’s first token to the network round trip. The caller measures one: how long the line stays quiet after they stop talking.
That gap is made of several pieces. First the system has to decide the caller has finished, and this wait is often the largest piece. Then recognition settles the last words, the language model starts its reply, and the voice produces its first audio. Finally the audio crosses the network to the phone and fills a small playback buffer.
People notice. Conversation research across ten languages found that speakers everywhere keep the silence between turns short. A reply that lands well after that sounds hesitant, and a long one makes the caller wonder if the line dropped.
Phone networks add their own delay on top, in both directions. The ITU’s G.114 recommendation exists because one-way delay hurts conversation as it grows, and mobile and internet calls often carry more of it than old landlines did.
Measure it the way the caller hears it. Record both sides of a real call, on a real phone, and time from the end of the caller’s speech to the first sound of the reply. Component numbers help you find where the time went. They don’t add up neatly to the total, because some pieces overlap and some sit outside your servers.
Then look at the slow end as well as the average. A median that feels fine can hide a slow tail, and the slow replies are the ones people remember.
Speed isn’t everything, though. An agent that answers instantly but cuts people off sounds worse than one that waits a beat for a number to finish.
In Voice Infra
This is what Voice Infra does about it.
Every call runs on our own GPUs in India, with speech recognition, the model and the voice side by side. Our agents reach a first reply in a median of 900 ms, though real calls vary with the caller's network and how they speak. On the Realtime API, every turn reports its own voice-to-voice time, so you can watch our side of the wait.