Low Latency Voice AI Benchmarks: How to Read and Test Them

Low latency voice AI is an agent that replies quickly enough that callers do not notice a gap, and the benchmark that matters is the time from the moment a caller stops speaking to the moment the agent starts speaking, measured on real phone calls. Two public benchmarks of live calls in 2026 found typical turn times (median or mean) between 1.27 and 3.76 seconds across platforms, far above the roughly 200 ms gap people leave in ordinary conversation. Because results depend so heavily on setup, measure latency on your own calls before you trust any single number.

What is a good latency for a voice AI agent?

No standards body sets a latency threshold for AI voice agents, so use two reference points and then set your own limit.

The first is human conversation. A study of 10 languages published in PNAS found a mean gap of 208 ms between one speaker finishing and the next starting, with the most common gap in every language falling between 0 and 200 ms (Stivers et al., 2009). None of the platforms in the benchmarks below comes close to that, which is why turn-taking design matters as much as raw speed. Our explainer on how voice AI latency affects the call experience covers what callers notice.

The second is vendor guidance, which is measured on the vendor's side of the call. Retell AI's docs say its platform reaches latency as low as 600 ms from the end of the caller's speech to the start of the agent's response, recommend keeping an agent's estimated latency under 1.5 seconds, and treat a median above 2.5 seconds on real calls as high.

What do published voice AI latency benchmarks show?

Independent, cross-platform latency data is scarce. The most complete public source is Cekura, a voice agent testing company that publishes its results and methodology. Its two 2026 platform boards measured on live calls are summarized below with two reference points.

• Cekura production agents, 2026: What was measured: 9 live agents, 270 phone calls, median. Turn time: 2.38s to 3.76s.
• Cekura Benchmark v1, Oct 2026: What was measured: 8 configurations, 246 calls each, mean. Turn time: 1.27s to 3.08s.
• Retell docs (vendor-reported): What was measured: Server side, excludes caller network. Turn time: As low as 600 ms.
• Human conversation (PNAS, 2009): What was measured: 10 languages, mean gap between turns. Turn time: 208 ms.

The two Cekura boards test different things. The production voice agents benchmark called nine real agents on their public phone lines with the same ten scenarios, three times each, for 270 calls, and reports median turn latency. Retell had the lowest median at 2.38 seconds, and the slowest was 3.76 seconds. Cekura notes that each agent serves a different business, so the ranking is indicative rather than definitive.

The Benchmark v1 leaderboard, last updated October 12, 2026, gave eight provider-chosen configurations the same agent brief and 82 scenarios, with 246 retained calls each, and reports mean response time. There, the fastest configuration averaged 1.27 seconds and the slowest 3.08 seconds, with Retell at 2.21 seconds. The same platform can rank first on one board and in the middle on another, which is the main lesson: setup decides the number.

Why do voice AI latency numbers disagree?

Most disagreement comes from four measurement choices. Check all four before comparing two numbers.

• Where it is measured. Retell's end-to-end metric runs from the end of the caller's speech to the start of the agent's speech on Retell's servers, and does not include the network trip to the caller. A phone-line benchmark measures from the caller's side, so it adds carrier and telephony time.
• Which statistic. A mean is pulled up by a few slow turns. A median (P50) shows the typical turn. P90 shows how slow one turn in ten is.
• What counts as a turn. Retell's metric counts only turns where the agent speaks, so time spent on silent tool calls does not inflate it. Another vendor may count differently.
• Which configuration. Model, voice, prompt length, and transcription settings all move latency. In Cekura's v1 board, each provider chose its own configuration.

What makes up end-to-end latency?

End-to-end latency is the sum of every step between the caller's last word and the agent's first sound.

End-to-end latency budget: End-to-end latency = Endpointing wait + Transcription + LLM time to first token + Text-to-speech first audio + Network and telephony

300 ms endpointing and transcription + 400 ms LLM + 150 ms text-to-speech + 150 ms network and telephony. 300 + 400 + 150 + 150 = 1,000 ms, so the LLM is 40% of this agent's delay.

Fix the largest component first, which you can only see in a per-call breakdown.

The endpointing wait is the step most teams overlook. After the caller stops, the system waits for a short silence to confirm they have finished. A longer wait produces more complete transcripts but adds directly to the delay the caller hears, and a shorter one risks cutting people off. Our glossary entry on turn-taking endpoints explains the mechanics, and our guide to interruptions and natural turn-taking covers the other half of the problem: how fast the agent stops when the caller talks over it.

How do you measure voice AI latency on your own pilot?

Measure from both sides of the call, on the same scenarios, and compare the results.

1. Fix the test conditions. Use the same scenarios, prompts, and agent version for every run, and place calls at the hours you expect real traffic.
2. Measure what the caller hears. Record a set of test phone calls. For a sample of turns, mark the end of the caller's speech and the start of the agent's reply in the audio and note the gap.
3. Pull the platform breakdown. In Retell, Call History shows per-call latency, and the Get Call API returns a latency object with end-to-end, transcription, LLM, text-to-speech, and knowledge base timings, each with P50, P90, P95, and P99 values. See how to check actual latency.
4. Compare the two. The difference between caller-side and platform-side numbers is your network and telephony overhead.
5. Test at your peak. Run a batch of calls at your expected peak concurrency, not just one call at a time.
6. Keep measuring after launch. Add a latency threshold to your QA criteria and an alert rule, so a slowdown shows up as an alert rather than as complaints.

Set your latency limit on P90, not the average. An average hides the slow turns. P90 shows how slow one turn in ten is, and slow turns are when callers start talking over the agent.

Which settings trade latency for accuracy?

Most latency fixes cost something else, so change one setting at a time and rerun your quality tests after each change. Retell's latency troubleshooting guide lists these levers:

• Model choice. Smaller models respond faster but reason less. Switching models changes behavior, so retest.
• Prompt length. Every token adds time on every turn, and prompts beyond roughly 8,000 tokens become noticeably slower.
• Fast tier. Routes model calls through dedicated infrastructure to cut average latency and variance, at a higher per-minute rate.
• Transcription mode. "Optimize for speed" responds sooner, while "optimize for accuracy" helps when capturing account numbers or dates. You can set accuracy on just the step that collects them.
• Response wait time. This deliberate pause before each reply can add up to 5.5 seconds at its most patient setting.
• Geography. Calls across continents add distance, so use a phone number in your callers' region.

Is the fastest voice agent the best one?

Not on its own. In Cekura's v1 board, the fastest configuration ranked third on repeatable reliability (scenarios passed on all three runs), and the configuration ranked first on reliability sat in the middle on speed. A fast agent that skips a required step or books the wrong slot is a failed call. If you are comparing platforms for inbound call center solutions, score latency next to resolution, accuracy, and transfers on one sheet, such as our voice agent QA scorecard, and follow the pilot process in our guide to evaluating a voice agent platform.

The only latency benchmark that settles a buying decision is the one measured on your own calls, from your callers' side. Retell runs a pilot on real calls before any contract, with per-call latency breakdowns your team can inspect. Talk to the Retell team about a pilot.