How do I compare AI voice agent latency between vendors?
Compare latency by making the same test calls to each vendor and timing the gap between you finishing a sentence and the agent starting to speak, over at least ten calls. Ignore headline numbers. Around 1 to 1.5 seconds feels natural on phone calls; beyond two seconds, callers start talking over the agent.
Why vendor numbers are hard to compare
Vendors quote latency differently. Some measure only the language model, some only the speech engine, and some the full round trip from the caller's last word to the agent's first sound. Only the full round trip, on a real phone line, matches what your customer experiences.
A simple test you can run
- 01Pick five questions of varying difficulty, such as a price, a location and an objection.
- 02Call each vendor's agent from a normal mobile phone, not a laptop microphone.
- 03Time the gap from the end of your sentence to the first sound of the reply, using a phone stopwatch or screen recording.
- 04Repeat at different times of day, including a busy evening.
- 05Record all calls and note the average and the worst case.
What to measure
| Metric | Why it matters |
|---|---|
| Average gap | Overall feel of the conversation |
| Slowest reply | Long pauses are what callers remember |
| Reply after a long answer | Complex queries can add delay |
| Interruption handling | Whether the agent stops when you cut in |
| Behaviour on a weak mobile network | Real customers are not on Wi-Fi |
Things that change the result
- —Your test language: some languages are slower than others on some systems.
- —Telephony path: the carrier and routing add delay of their own.
- —Voice tier: richer voices can take slightly longer to produce.
- —Script length: a long, multi-part reply takes longer to start than a short one.
What latency is acceptable
Human conversation leaves gaps of a fraction of a second. Phone agents cannot match that, but a gap of roughly one to one and a half seconds feels acceptable to most callers. Beyond two seconds callers start repeating themselves or talking over the agent, and conversations degrade quickly.
Outpero's figure
Outpero's end-to-end conversational latency is 0.8 to 1.3 seconds on average, measured from the caller stopping to the agent speaking. Treat this as a vendor claim and measure it yourself on your own calls. Outpero publishes this page, and measuring several vendors the same way is the only fair test.
Want to hear an AI voice agent handle this on a real call?
Follow-up questions
Is the lowest latency always best?
+
Not always. An agent that replies too fast can cut in before the caller has finished, especially with pauses or Hinglish. Good turn-taking, which detects when a caller has really finished, matters as much as raw speed. Judge by how natural the whole conversation feels.