Time to first byte
In one sentence
Time to first byte (TTFB) in voice AI is how quickly a speech synthesizer starts emitting audio after it receives text, which is the moment the listener's wait ends, and it shares its name with an unrelated web-performance metric that has to be qualified apart from it.
Definition
Time to first byte is how long until the first piece of data arrives. In voice work it usually means how quickly a speech synthesizer starts emitting audio after it receives text.
What matters is not how long the whole answer takes to produce, but how soon the sound begins, because that is when the waiting stops.
The term comes from web performance, where it measures the delay before a server sends its first response byte. Voice borrowed it for the synthesis stage specifically, and the two usages now sit side by side, which is where the confusion starts.
Synthesis TTFB, the voice usage
- It measures from text arriving at the synthesizer to the first audio byte coming out.
- Streaming synthesis reaches roughly 100 to 200 milliseconds. Non-streaming synthesis has no meaningful time to first byte, because nothing emerges until the whole thing is generated.
- That is why streaming synthesis is a requirement for conversation, not an optimization you add later.
Why it matters more than total synthesis time
- Total synthesis time for a long answer can run into seconds. If playback begins after 150 milliseconds and keeps going as generation proceeds, the listener never experiences that total.
- The pipeline should start synthesizing the first clause while the language model is still writing the rest. That overlap is where most of the perceived latency saving in a voice pipeline comes from.
Page TTFB, the original usage
- It still matters to a voice widget, because the loader script and the page it sits on have their own performance to answer for.
- A widget that delays page rendering harms Core Web Vitals. That is a separate concern from conversational latency, and it is what the web team who has to approve the install will look at.
The measurement ambiguity
- Because both usages show up in voice projects, TTFB should always be qualified. Synthesis TTFB and page TTFB are unrelated numbers that happen to share a name.
- On a spec sheet, assume the synthesis meaning unless page performance is the subject under discussion.
Common misconception
That fast total synthesis matters more than a fast first byte. For a listener, the wait ends when sound begins. Total generation time is invisible as long as playback keeps ahead of it.
Why it matters commercially
Synthesis time to first byte is a small, controllable component that decides whether an agent feels responsive, and it is one of the easier numbers to verify in a vendor evaluation. Because the name is shared, resolving which TTFB is meant is worth doing up front in any conversation that spans both voice and web.
In voice specifically
A page of text is judged on whether it is right, and it can render all at once. Spoken audio has to arrive in order and in time, so the moment the first byte plays is the moment the interaction stops feeling like a wait. There is no on-screen partial answer to cover the gap, which is why first-byte latency carries more weight in speech than a total generation time ever does.
Where AsqVox fits
AsqVox installs as a single script tag, so the first thing a web team checks is what that tag does to page load, which is the page-performance sense of time to first byte and a separate question from how quickly the voice answers once a visitor starts talking. Both have to be good: one gets the widget approved, the other keeps the conversation feeling live.
Visual
When the waiting actually stops
The listener's wait ends when sound starts, not when generation finishes.
Statistics
Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.
Streaming synthesis time to first byte typically lands at 100 to 200 milliseconds. Non-streaming synthesis has no meaningful first-byte figure, because nothing is emitted until generation completes.
100 to 200 msindustry rangeCommon production budgeting convention, 2026 - An engineering convention drawn from production practice rather than a published research finding. It applies to streaming synthesis only.
Sub-second time to first audio is the working design target for conversational voice.
Under 1,000 msindustry rangeIndustry working threshold, 2026 - A working convention, not a published standard. Synthesis time to first byte is one segment inside this budget.
Past roughly 1,500 milliseconds, a conversation is widely reported to feel broken.
1,500 msindustry rangeIndustry working threshold, 2026 - A shared working line rather than a measured cliff.
Human conversational turn gaps average around 200 milliseconds, which places synthesis time to first byte at roughly the scale of a natural pause.
200 msindustry rangeWidely cited conversational turn-taking baseline, 2026 - A reference point, not a target. It is useful here because it shows a good first-byte figure is about the length of the pause a person would leave anyway.
A voice widget loads on every page view, whether or not the widget is ever used, so its contribution to page performance is a separate and legitimate concern for web teams. Core Web Vitals impact is a standard question in widget evaluation.
Every page viewindustry rangeStandard widget evaluation practice, 2026 - This is the page-TTFB half of the name, unrelated to synthesis timing. It is a real approval gate, not a latency number.
There is no published independent benchmark comparing synthesis time to first byte across text-to-speech providers under matched conditions.
-no reliable figureSo a first-byte figure from a vendor is worth verifying yourself against the response lengths and network your visitors actually use, not taking on the published number.
Examples
In practice
A team measures total synthesis time at 2.4 seconds for a long answer and concludes synthesis is the bottleneck. Instrumenting first-byte emission shows 160 milliseconds, with playback comfortably ahead of generation the whole way through. The listener never waited 2.4 seconds. The metric being optimized was the wrong one.
The everyday version
Time to first byte is how quickly the voice starts talking, as opposed to how long the whole answer takes to produce. As long as it starts quickly and keeps up, nobody notices or cares how long the full sentence took to generate.
Usage
Who says it
- Web performance engineers, meaning the original page metric.
- Voice engineers, meaning synthesis first-byte latency.
- The ambiguity is real, and worth resolving explicitly in any conversation that touches both.
Where it turns up
- Next to streaming synthesis support and latency figures, and separately, next to page load impact.
- A spec sheet that lists a TTFB without saying which one is inviting a comparison of two different measurements.
Common misuse
- Failing to qualify which TTFB is meant.
- Optimizing total synthesis time rather than first-byte time.
- Treating streaming synthesis as an enhancement rather than a requirement for conversation.
Questions people ask
What is time to first byte in voice AI?
It is how quickly a speech synthesizer starts emitting audio after it receives text, measured from text arriving at the synthesizer to the first audio byte coming out. Streaming synthesis reaches roughly 100 to 200 milliseconds. It is the moment the listener's wait ends, which is why it matters more than how long the whole answer takes to generate.
What is the difference between synthesis TTFB and page TTFB?
They are two unrelated numbers that share a name. Synthesis time to first byte measures text in to first audio out, and belongs to conversational latency. Page time to first byte is the original web-performance metric, measuring the delay before a page server sends its first response byte, and it governs Core Web Vitals. On a spec sheet, assume the synthesis meaning unless page performance is the subject.
Why does total synthesis time not matter to the listener?
Because the wait ends when sound begins. If playback starts after about 150 milliseconds and keeps ahead of generation, the listener never experiences the full generation time, even when it runs into seconds. In one case a 2.4-second total synthesis had a 160-millisecond first byte, and the listener never waited 2.4 seconds. First-byte time is the number to optimize, not total time.
Is streaming synthesis required for a voice agent?
For conversation, yes. Non-streaming synthesis emits nothing until the whole answer is generated, so it has no meaningful time to first byte and the listener waits for the entire generation. Streaming synthesis emits the first audio byte at roughly 100 to 200 milliseconds and lets playback run alongside continuing generation, which is why it is treated as a requirement rather than an optimization.
Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.