Lowest-Latency AI Voice Agents: How to Test Before You Buy

What latency means on a phone call

Latency in a voice agent is the silence between the caller finishing a sentence and the agent starting to reply. It is the single biggest reason a call feels natural or robotic, and it is measured in the caller's ears, not on a server dashboard. The short definition is in our glossary entry on latency.

It is not the same thing as speed to lead. Calling a new enquiry within 30 seconds of the form being submitted is a different clock; this guide is only about the gap inside the conversation.

People set the benchmark without noticing. A 2009 study in the Proceedings of the National Academy of Sciences by Stivers and colleagues, covering ten languages, found that the typical gap between turns in human conversation is around 200 milliseconds. No voice agent reaches that, so the practical question is how far from natural a call can drift before a caller notices and starts to talk over the agent, repeat themselves or say hello twice.

The network alone uses part of the budget. The ITU's G.114 recommendation treats one-way delay under about 150 milliseconds as fine for most conversations, and a mobile call in a congested area can spend a good share of that before any AI software has run.

The latency budget: where the time goes

Every turn of a call passes through the same stages. Vendors differ in how they build each one, but the stages are the same, and the delay you hear is the sum of all of them. A vendor who quotes only one stage is telling the truth about a small part of the call.

Two things follow. First, a low number for one stage does not guarantee a fast call. Second, the stages that you control as a business are the ones nobody advertises: how long your API takes, how long you let the agent talk, and which language mix your callers use.

What vendors publish, and why it is hard to compare

Here is every latency figure the vendors in our dataset publish, with where it appears. All are the vendor's own claims.

On paper, Exotel, Bland, Synthflow and Retell publish lower figures than Outpero's 0.8 to 1.3 seconds, and for a business with engineers who want to tune every layer, those platforms are worth testing. The catch is definition. A figure measured at the model, on a server, in English, says little about a Hindi call on 4G in Pune.

Four questions to ask about any latency claim

1. 01Where do the start and end points sit? Caller stops speaking to agent's first audio is what a caller hears. Model response time is only one stage.
2. 02Which percentile is it? An average hides the slow calls. The worst one in ten is what makes customers hang up.
3. 03Which language and which route? A figure measured in English over a data-centre connection is not the figure on a Telugu call over mobile.
4. 04With or without tools? A booking that checks a calendar takes longer than a greeting. Ask for the figure on a call that does real work.

What moves latency on Indian calls

• —Language mix. Callers in Hyderabad, Chennai or Pune rarely stay in one language. The agent has to recognise Hinglish, Tenglish or Tanglish without a pause that looks like confusion.
• —Call route. Where the audio is carried matters. Ask where the telephony is hosted and whether an Indian call is routed through India.
• —Reply length. A two-sentence answer that starts speaking at once feels faster than a perfect paragraph that starts late. Keep instructions short and ask the agent to answer first, elaborate on request.
• —Live lookups. Each calendar, sheet or API call adds time inside the turn. Test with the lookups switched on, not off.
• —Voice tier. Voices differ in how they are built and delivered. Outpero's Premium tier is the ultra-low-latency option with the most lifelike voices; Standard is the most chosen for everyday sales and support.

The stopwatch test: measure any voice agent yourself

You do not need a lab. You need ten minutes, two phones and a script. Run the same test on every vendor you shortlist, from the same place, at the same time of day.

1. 01Write ten caller lines in your actual language mix. Three short (a yes, a name), three medium (a question about price or timing), two code-switched, two that need a lookup, such as a booking. A realistic opener in Hindi: "मुझे कल शाम का appointment चाहिए, कितने बजे free हैं?" (I need an appointment tomorrow evening, what times are free?). In Telugu: "రేపు సాయంత్రం appointment ఉందా?" (Is there an appointment tomorrow evening?)
2. 02Call from an ordinary mobile on mobile data, from the place your customers call from, not from office Wi-Fi.
3. 03Record the call on a second phone placed near the speaker, or use the platform's own recording. You need audio you can play back.
4. 04Start the stopwatch when you say your last syllable and stop it at the agent's first syllable. By hand, expect an error of roughly a quarter of a second, which is enough to tell 1 second from 3. For precision, drop the recording into a free audio editor such as Audacity; the gap between the two waveforms is exact.
5. 05Do ten calls, at different times of day. Write down the median and the worst call. The median is the typical experience; the worst is what your angriest customer gets.
6. 06Interrupt the agent mid-sentence three times and time how long it takes to stop talking. A voice agent that keeps going over you is worse than a slow one.
7. 07Run the lookup lines and note any silence after the agent says it is checking. A long gap with no filler is where callers hang up.

Reading the result

These bands are our own rule of thumb, not an industry standard. The point is to compare vendors on the same test, not to chase a magic number.

How Outpero handles latency

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. The Premium tier, at ₹6.8 per connected minute, is the ultra-low-latency option and carries the most lifelike voices, meant for calls where the voice is the brand. Standard, at ₹5 per connected minute, is the most chosen tier for everyday sales and support. Both tiers cost the same in every language, and both prices exclude 18% GST.

The 0.8-to-1.3-second average is a product-level number, so time the tier you intend to buy. A new account gets 50 free credits with no card, enough for the whole ten-call test above on either tier. There is no setup fee and no subscription; a dedicated +91 number is ₹649 a month and includes one channel.

The agent also switches to English the moment the caller does, mid-sentence and in the same voice, so the code-switched lines in your script are a fair test of it. Outpero calls a new lead within 30 seconds, answers inbound calls on your business number, and can transfer the call to your team when a person should take it.

Where Outpero is not the right choice on latency

If you need a sub-500-millisecond claim and have engineers to tune it, test a developer platform such as Retell or Bland alongside Outpero. You will then assemble the telephony, the compliance and the language handling yourself, and Vapi's and Bland's own pricing pages show model or telephony costs billed on top of the platform fee. If you need a language Outpero does not support, such as Bengali or Gujarati, latency is not your first problem.

Faster is not always better

An agent tuned to answer instantly can cut callers off. Indian callers often pause mid-sentence to find the English word for something, and an agent that treats the pause as the end of the turn will talk over them. An agent that waits for the caller is better than a 0.8-second agent that interrupts.

Latency also matters unevenly by job. It matters most in sales qualification, appointment booking and receptionist calls, where a conversation is the product. It matters little for voice broadcast, which plays one fixed message and has no turns at all, and moderately for reminders, which are short and predictable.

Frequently asked questions