Komms · Latency benchmark
Voice AI latency, measured on live calls: Komms TTFW is a median 77 ms
- $0.08 a minute, phone call included
- $15/mo per phone number
On live phone calls, a Komms voice agent hands its first words to the voice a median of 77 milliseconds after the caller's words arrive. About 97% of replies start within 200 ms, and 99.9% within 500 ms. These numbers come from the call logs of 2,841 real calls, 15,297 caller turns, between October 21 and November 6, 2026. From October 14 to 20, across 6,415 turns, the median was 98 ms. Below: every number, how we measured it, and exactly what the clock covers.
Heads up: the demo line is a live Social Security disability (SSDI) intake. It asks real screening questions and calls are recorded. Hang up any time.
The numbers
Time is counted per caller turn, from the moment the caller's words arrive as text. All times are in milliseconds (1,000 ms = 1 second). Calls from October 21 to November 6, 2026.
| What we timed | Median | 90% within | 95% within | Under 200 ms |
|---|---|---|---|---|
| First words handed to the voice (time to first word, TTFW) | 77 ms | 121 ms | 153 ms | 96.8% |
| The full answer handed to the voice | 89 ms | 701 ms | 968 ms | 70.5% |
Over time
The same clock on the same campaign, by period.
| Period | Caller turns | Median | 90% within | 95% within |
|---|---|---|---|---|
| Early October 2026 | — | 162 ms | — | — |
| October 14 to 20 | 6,415 | 98 ms | 158 ms | 177 ms |
| October 21 to November 6 | 15,297 | 77 ms | 121 ms | 153 ms |
- TTFW
- About 97% of turns started within 200 ms, and 99.9% within 500 ms.
- Full answer
- 71% of answers were handed to the voice within 200 ms, 82% within 500 ms and 96% within one second.
- Faster than three weeks ago
- In early October the median on this campaign was 162 ms. The daily median went from 139 ms on October 14 to 78 ms on October 20, and was 77 ms on October 21 and 22.
- Short bridge phrases
- On 13.9% of turns the agent opened with a short phrase such as "let me check" while it worked out the answer. On those turns the full answer followed a median of 614 ms after the caller's words arrived. Those turns are counted in every row above.
Where the time in a call goes
We added up the deciding time on every one of the 621 calls from October 14 to 20. A typical call ran about two minutes, with 11 caller turns.
| Part of the call | Share of call time | Typical call |
|---|---|---|
| Komms deciding what to say | 1.7% | 1.7 seconds in total |
| Everything else: Komms speaking, the caller thinking and speaking, the transfer | 98.3% | about 2 minutes |
Across a whole call, the agent spends under two seconds working out its replies. The rest is the conversation itself.
How we measured
- Real calls
- Live phone calls on three campaigns, inbound calls and callbacks: 2,841 calls with a timed reply between October 21 and November 6, 2026, and 621 calls on one campaign from October 14 to 20. Each call ran the full script and ended with a transfer to a person. Median call length: about two minutes. No test calls and no lab setup.
- Every turn counted
- All 15,297 caller turns from October 21 to November 6 (and all 6,415 from October 14 to 20) are in the numbers, including the slowest. Nothing was trimmed. Each campaign on its own had a median between 76 and 78 ms.
- Where the clock starts
- When the caller's words arrive from speech-to-text as text. Komms writes that moment to the call log.
- Where the clock stops
- When Komms hands the first words of its reply to the text-to-speech voice. Komms writes that moment to the call log too.
- Where the numbers come from
- The Komms call log. Every call report shows the timing of every turn, so any Komms customer can check the same numbers on their own calls.
What time to first word covers
A voice reply has four steps. Time to first word is steps 2 and 3: the part the voice agent platform controls.
| Step | What happens | In our 77 ms? |
|---|---|---|
| 1. Hear | Speech-to-text notices the caller has finished and turns the speech into text. | Before the clock starts |
| 2. Decide | The agent works out what to say. | Yes |
| 3. Hand off | The first words of the reply go to the voice. | Yes |
| 4. Speak | The voice turns the words into sound and the phone network carries it to the caller. | After the clock stops |
What the caller hears is all four steps added together. Steps 1 and 4 add time on every voice platform, so the pause a caller hears is longer than 77 ms.
Three terms that get mixed up. Time to first word (TTFW) is steps 2 and 3. Time to first audio is how long before the first sound plays: for a text-to-speech voice alone, a slice of step 4, or for the whole reply, all four steps. End-to-end or voice-to-voice latency is all four steps. A number is only comparable with another number that uses the same clock. What is time to first word? · What is time to first audio?
What voice AI platforms publish about their own latency
Each figure below is the vendor's own, read on the vendor's own page on October 21, 2026. Vendors start and stop the clock in different places, so we print what each one says it measures and leave the comparison to you.
Voice agent platforms
| Platform | Published figure | What it says it measures | Source |
|---|---|---|---|
| Komms (KX21) | 77 ms median, 153 ms at the 95th percentile | Caller's words arrive as text → first words handed to the voice. Live phone calls, 15,297 turns, October 21–November 6, 2026. | This page |
| Twilio ConversationRelay | 491 ms median, 713 ms at the 95th percentile | "Internal benchmarks … using different models." Start and end points not defined on the page. | twilio.com |
| Retell AI | ~600 ms | "~600ms latency." The page does not say median or average. | retellai.com |
| Vapi | under 500 ms average; "sub-600ms response times" | "<500ms average latency" on the home page; "Sub-600ms response times" in the docs. | vapi.ai, docs |
| Synthflow | under 500 ms | "sub-500ms latency." No breakdown given. A separate "sub-100 ms" figure on the page is for its own phone network. | synthflow.ai |
| PolyAI | under 300 ms | "<300ms model response, voice to voice." | poly.ai |
| Bland AI | under 200 ms | "End-to-end response runs under 200ms on the standard model tier." | bland.ai |
| Telnyx Voice AI | under 200 ms | "Sub-200ms latency," credited to "co-located GPUs and telephony PoPs." The page does not say where the clock starts or stops. | telnyx.com |
Text-to-speech voices (one step inside a voice agent)
| Voice | Published figure | What it says it measures | Source |
|---|---|---|---|
| ElevenLabs Flash v2.5 | ~75 ms | "Ultra-low latency (~75ms)", footnoted as excluding application and network latency. | elevenlabs.io |
| Cartesia Sonic | under 90 ms | "First audio in under 90ms." | cartesia.ai |
| Deepgram Aura-2 | under 200 ms | "Sub-200ms time-to-first-byte." | deepgram.com |
The Komms row uses the time-to-first-word clock described above. The other rows use each vendor's own clock (fastest voice AI, compared). Where a vendor says "end-to-end" or "voice to voice", its figure includes steps that our 77 ms leaves out.
Check it yourself
Watch the short version. What is TTFW? (52 seconds)
Run it on your own calls. Komms is $0.08 a minute with the phone call, the AI and a standard voice included, and a phone number is $15 a month. New accounts start with 625 free minutes, with no credit card. Every call you run shows the same per-turn timing in its call report.
Build your first agent in about 10 minutes → Fastest voice AI, compared How Komms works Pricing
More Komms setups: AI receptionist · AI call center · AI outbound calling · AI answering service · AI voice agent · AI phone agent
Questions
- What is time to first word (TTFW) in voice AI?
- Time to first word is how long a voice agent takes to start its reply: from the moment the caller's words arrive as text to the moment the first words of the reply are handed to the voice. It covers the agent deciding what to say. It does not include speech-to-text noticing the caller has finished, or the voice and phone network delivering the sound.
- How fast does a Komms voice agent reply?
- On 15,297 caller turns from 2,841 live phone calls between October 21 and November 6, 2026, Komms handed its first words to the voice a median of 77 milliseconds after the caller's words arrived. 90% of turns were within 121 ms and 95% within 153 ms. From October 14 to 20, the median was 98 ms, and in early October it was 162 ms.
- Is 77 ms the pause the caller hears?
- No. 77 ms is the time Komms takes to decide and start its reply. The caller also waits for speech-to-text to notice they have finished and for the voice and the phone network to deliver the sound. Those steps add time on every voice platform.
- What is the difference between TTFW, time to first audio and end-to-end latency?
- Time to first word is the agent deciding what to say and handing it to the voice. Time to first audio is how long a text-to-speech voice takes to make its first sound. End-to-end or voice-to-voice latency is everything from the caller finishing to the caller hearing the reply. Only numbers that use the same clock can be compared.
- How much of a call does the AI spend thinking?
- On the 621 live calls we measured from October 14 to 20, Komms spent 1.7% of the call deciding what to say: about 1.7 seconds in total across a typical 11-turn, two-minute call. The rest is Komms speaking, the caller thinking and speaking, and the transfer.
- Can I check these numbers on my own calls?
- Yes. Every Komms call report shows the timing of each turn, taken from the same call log these numbers come from. New accounts start with 625 free minutes, with no credit card.