What is a good latency for an AI voice agent?
A good latency for an AI voice agent is under about 1.5 seconds, measured from the moment the caller stops speaking to the moment the agent starts replying. Below that, a phone conversation feels natural; beyond two seconds, callers start repeating themselves or saying 'hello?'. Outpero's agents run at 0.8 to 1.3 seconds on average end to end, measured exactly that way.
What latency actually means on a call
Voice-agent latency is the silence the caller hears between finishing a sentence and hearing the first word of the reply. It is the sum of several steps: deciding the caller has finished, finalising the transcript, generating the start of the reply, synthesising the first audio and pushing it down the phone network.
Vendors sometimes quote only one of these, such as model response time, which can look impressive while the real gap on a call is much longer. The only number that matters to the caller is end to end: caller stops, agent speaks.
How it compares with human conversation
Research on human turn-taking, including a 2009 PNAS study by Tanya Stivers and colleagues across ten languages, found that people typically start replying within a few hundred milliseconds of the other person finishing. Machines cannot quite match that over a phone network yet, but the closer they get, the less the caller notices.
| End-to-end gap | How it tends to feel |
|---|---|
| Under 1 second | Very natural, close to a human |
| 1 to 1.5 seconds | Natural for a phone call; a short thinking pause |
| 1.5 to 2.5 seconds | Noticeable; some callers start talking again |
| Over 2.5 seconds | Feels broken; callers say 'hello?' or hang up |
These bands are a practical rule of thumb, not an industry standard.
What pushes latency up
- —Waiting too long to decide the caller has finished speaking.
- —Using a large, slow model for every turn, even simple acknowledgements.
- —Generating the whole reply before starting to speak instead of streaming it.
- —Servers far from the caller, which add network round trips on every turn.
- —Calling slow external systems mid-turn without saying anything to fill the gap.
Where Outpero sits
Outpero measures latency the way the caller experiences it, and its agents run at 0.8 to 1.3 seconds on average end to end. The Premium voice tier at ₹6.8/min uses curated native-speaker voices with ultra-low latency, for calls where every pause counts. When you evaluate any vendor, place a real test call to your own phone and time the gaps yourself rather than relying on a demo video.
Want to hear an AI voice agent handle this on a real call?
Follow-up questions
Is lower latency always better?
+
Mostly, but not at any cost. An agent tuned to reply instantly can jump in while the caller is only pausing to think, which feels rude and causes overlapping speech. The best systems balance fast replies with patient turn detection, so they respond quickly to finished sentences and wait through mid-sentence pauses.