Learn · What is time to first word?
What is time to first word (TTFW) in voice AI?
Time to first word is how long a voice AI agent takes to start its reply after the caller finishes talking. It is the number that tells you whether a phone agent feels quick or makes people wait.
Komms on live calls: a reply starts in a median of 77 ms
Measured on 1,364 caller turns from 138 real phone calls, October 21 and 22, 2026, from the Komms call logs. The clock runs from the caller's words arriving as text to Komms handing the first words of its reply to the voice.
| Komms, live calls | Median | 90% within | 95% within | Under 200 ms |
|---|---|---|---|---|
| Reply starts (October 21–22, 2026) | 77 ms | 123 ms | 155 ms | 97% of turns |
| Same line, October 14–20 (6,415 turns) | 98 ms | 158 ms | 177 ms | 97% of turns |
| Same line, early October | 162 ms | — | — | — |
- Real calls, every turn
- A live benefits-screening line. No test calls, no lab, and the slowest turns are counted.
- Checkable
- Every Komms call report shows the timing of every turn, so customers see the same numbers on their own calls.
- What it leaves out
- Speech-to-text before the clock starts, and the voice and phone line after it stops. Those add time on every platform.
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.
Full numbers, method, and what other platforms publish Watch: What is TTFW? (52 seconds)
The four steps of every reply
Each time a caller stops talking, a voice agent goes through four steps before the caller hears an answer.
| Step | What happens |
|---|---|
| 1. Hear | Speech-to-text notices the caller has stopped talking and turns their words into text. |
| 2. Decide | The AI works out what to say. |
| 3. Hand off | The first words of the reply go to the voice. |
| 4. Speak | The voice turns the words into sound, and the phone network carries it to the caller. |
Time to first word is about steps 2 and 3: how fast the agent decides and starts its reply. Steps 1 and 4 depend mostly on the speech-to-text service, the voice and the phone line.
Two ways people use the term
| Clock | Starts | Stops | Used by |
|---|---|---|---|
| Platform clock | The caller's words arrive as text | The first words of the reply go to the voice | Voice agent platforms reading their own call logs, including Komms |
| Caller clock | The caller stops talking | The caller hears the first word | Testing tools that record calls from the outside |
The caller clock is always the longer of the two, because it includes all four steps. Before you compare two numbers, ask where each clock starts and where it stops. A figure without that answer cannot be compared with anything.
Time to first word vs time to first audio vs end-to-end
| Term | What it covers |
|---|---|
| Time to first token (TTFT) | How fast the AI model writes the first piece of its answer. One part of step 2. |
| Time to first word (TTFW) | The agent deciding and starting its reply: steps 2 and 3, or all four steps on the caller clock. |
| Time to first audio (TTFA) | How long before the first sound plays. Used for the voice alone (part of step 4) or for the whole reply. What is time to first audio? |
| End-to-end or voice-to-voice latency | All four steps: from the caller stopping to the caller hearing the reply. |
What is a good number?
In everyday conversation, people usually answer each other within a fraction of a second. A gap of a second or more starts to feel slow, and callers begin talking over the agent or asking "hello?".
The agent's own part, steps 2 and 3, is the part a platform can make small. When it takes a few hundred milliseconds or less, most of the pause the caller hears comes from speech-to-text, the voice and the phone line.
How to measure it fairly
- Use real calls
- Lab tests on short questions run faster than live calls with real callers.
- Report the median and the slow end
- The median (the middle turn) shows a typical reply. The 90th and 95th percentiles show how slow the slow replies get.
- Count every turn
- Leaving out the slowest turns makes any number look better.
- Watch for bridge phrases
- A quick "let me check" starts the reply sooner. Time the real answer separately too.
Start free Call it now: +1 (618) 228-1183
50,000 free tokens for new Komms accounts — about 625 minutes. No credit card needed.
Questions
- What does time to first word (TTFW) mean?
- It is how long a voice AI agent takes to start its reply after the caller finishes talking. Platforms usually time it from the caller's words arriving as text to the first words of the reply going to the voice. Testing tools often time it from the caller stopping to the caller hearing the first word.
- What is a good time to first word for a voice agent?
- The agent's own part, deciding and starting the reply, can be a few hundred milliseconds or less. The whole pause the caller hears should stay under about a second, because longer gaps make callers talk over the agent.
- Is time to first word the same as latency?
- It is one kind of latency. End-to-end or voice-to-voice latency covers everything from the caller stopping to the caller hearing the reply. Time to first word on the platform clock covers only the agent deciding and starting its reply.
- What is the difference between time to first word and time to first token?
- Time to first token is how fast the AI model writes the first piece of its answer. Time to first word includes that, plus everything else the agent does before its reply goes to the voice.
- What is Komms' time to first word?
- On 1,364 caller turns from 138 live phone calls on October 21 and 22, 2026, Komms handed its first words to the voice a median of 77 milliseconds after the caller's words arrived as text, with 95% of turns within 155 ms. That covers deciding and starting the reply only. The full method is on the Komms latency page.