Time to first audio
In one sentence
Time to first audio (TTFA) is the end-to-end wall-clock delay from the moment a person stops speaking to the moment the reply becomes audible, and it decides more than any other number whether a voice agent feels alive.
Not to be confused with time to first token.
Definition
Time to first audio measures the wait a visitor sits through after they stop speaking, ending the moment the answer becomes audible.
More than any other measurement, it separates an agent that feels alive from one that feels like a machine working something out.
TTFA belongs to the whole system rather than to any model inside it: wall-clock time, read end to end. Several components contribute, each varying on its own, and reading that budget correctly is most of the discipline.
The budget, component by component
- Endpointing. Deciding the speaker has actually finished, usually after 150 to 300 milliseconds of trailing silence. This is a straight tradeoff: a shorter threshold feels faster and cuts people off mid-thought.
- Transport upstream. Moving that last slice of audio across the network to the server.
- Transcription finalization. Committing the closing words of the utterance.
- Retrieval, if the system uses RAG. Turning the question into an embedding, running the search, reranking what comes back.
- Model time to first token. How long before the model emits anything at all.
- Speech synthesis time to first byte. Streaming text-to-speech usually lands in the 100 to 200 millisecond band.
- Transport downstream, and the browser starting playback.
Streaming is what separates a system that feels laggy from one that does not. Queue the stages and every one of them lands in the total, because nothing starts until the stage ahead of it has finished. Overlap them instead and synthesis can begin on the opening clause of the answer while the model is still composing the rest. That choice gets made early in the architecture and is expensive to undo.
Terms that sit next to it and are not it
- TTFT, time to first token, measures the model alone. It is one component of TTFA, and a vendor quoting TTFT as though it were TTFA is flattering itself.
- End-to-end latency is ambiguous. It can mean TTFA, and it can mean the point at which the whole response finishes. Always ask which one is on offer.
- Barge-in latency runs the other direction: once someone talks over the agent, how quickly the agent stops. Roughly 60 milliseconds is the working figure.
Common misconception
That faster is always better. Drop below roughly 200 milliseconds and a reply lands as an interruption instead of an answer, because no person could have weighed the question that quickly. Some production systems put a small delay back in on purpose, or cover the gap with a brief acknowledgment sound. The target is natural, not minimal.
Why it matters commercially
TTFA is the most honest demo metric in the category, and the hardest to game. Nobody can assess retrieval quality in a two-minute demo, and everybody can feel 1,500 milliseconds. It is also the number most likely to slip quietly in production as the corpus grows and retrieval slows, which makes it a monitoring commitment rather than a launch checkbox.
In voice specifically
In text you can cover a delay with a typing indicator or a partial answer already on screen. In speech there is nothing to show, so silence is the only signal the listener gets. That is why some systems fill the gap with a short sound instead of trying to shorten it.
Visual
Where the milliseconds go
Nobody can evaluate your retrieval in a demo. Everybody can feel a second and a half.
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.
The average gap between turns in human conversation is around 200 milliseconds. Every design target in voice AI is set against that number.
200 msindustry rangeWidely cited conversational turn-taking baseline, 2026 - A reference point rather than a target. It is quoted widely enough that no single study owns it, so cite it as the human baseline, not as a measurement of anything you built.
Landing under one second is the design target the industry generally works to.
Under 1,000 msindustry rangeIndustry working threshold, 2026 - A working convention, not a published standard. No body sets it and no vendor is audited against it.
Past roughly 1,500 milliseconds, a conversation is widely reported to feel broken.
1,500 msindustry rangeIndustry working threshold, 2026 - Same status as the sub-second target: a shared working line rather than a measured cliff. It is the number a buyer will feel in a demo without being told to look for it.
Synthesis should be cut within roughly 60 milliseconds of a person barging in.
60 msindustry rangeIndustry working threshold, 2026 - The mirror image of TTFA and a separate budget. A system can hit its TTFA target and still feel rude if it talks over an interruption.
Silero VAD scores a 30 millisecond audio chunk in under 1 millisecond, which confirms that voice activity detection is not normally the bottleneck.
Under 1 ms per 30 ms chunkvendor claimSilero VAD published benchmark, 2026 - Published by the project itself. The useful conclusion is directional: when TTFA is bad, detection is rarely where to look first.
Measured time to first token across leading realtime models: around 0.78 seconds for xAI Grok Voice, 0.82 for OpenAI gpt-realtime-1.5, 1.14 for Amazon Nova 2 Sonic, and 2.98 for Gemini 3.1 Flash Live.
0.78s to 2.98sindependentArtificial Analysis, via Softcery, April 2026, 2026 - Two cautions when citing these. The figures are TTFT rather than TTFA, so any complete pipeline sits above them. And the fastest and slowest of the leading models are separated by close to four times, which is enough for the choice of model on its own to settle whether a product comes in under a second.
Production budgets commonly allow 150 to 300 milliseconds for the endpointing silence threshold.
150 to 300 msindustry rangeCommon production budgeting convention, 2026 - An engineering convention rather than a published research finding, and it should be presented that way. Shortening it buys latency and starts cutting people off mid-sentence.
Production budgets commonly allow 100 to 200 milliseconds for speech synthesis time to first byte.
100 to 200 msindustry rangeCommon production budgeting convention, 2026 - Also an engineering convention rather than a research finding, and it applies to streaming synthesis only.
The component budget figures in circulation are engineering conventions, not published research findings. No independent study establishes them.
-no reliable figureUse them to plan a budget. Do not cite them as evidence, and do not let a vendor cite them at you as though they were.
No independent cross-vendor benchmark of end-to-end time to first audio exists. The public independent measurements cover time to first token only, which is one component of the budget.
-no reliable figureThis is why quoting TTFT as TTFA is both so common and so hard to check. The only reliable TTFA figure is one you measure yourself, on your own stack, over the network your visitors actually use.
Examples
In practice
TTFA comes back at 1,900 milliseconds and the team blames the model. Timing each stage separately tells a different story: 700 milliseconds inside the model, 800 inside retrieval. The vector search has no index, and the corpus it scans has grown fourfold since launch. With an index in place retrieval falls to 90 milliseconds and TTFA to 1,190. The model was never the problem.
The everyday version
You finish asking the voice agent something. If the reply arrives while you are still settling back into your chair, the exchange feels like one you would have with a person. If instead you get long enough to wonder whether it registered the question at all, you are waiting on a computer, and the tab is probably about to close.
Usage
Who says it
- Voice engineers live in this number and never quote it without percentiles, since a healthy median sitting on top of an ugly p95 still ships as a bad product.
- Sales engineers reach for TTFA in a demo because nothing else lets a buyer feel a difference that fast.
- Procurement meets TTFA as a service level commitment, and will ask which percentile is being promised, where the clock is read, and what happens if the number is missed.
Where it turns up
- Usually in the same clause as concurrency, uptime, geographic point of presence, tolerance for jitter and packet loss, and p50 p90 p95 p99 latency.
- A well-specified RFP says exactly where the clock starts and where it stops. An unspecified one produces bids that cannot be compared, because every vendor will measure from whichever point flatters it most.
Common misuse
- Quoting model TTFT as though it were system TTFA. No overclaim in the category is more common.
- Quoting a median and no percentile. What gets remembered about a voice product is the bad calls, not the average one.
- Measuring on a wired lab connection and presenting the result as production. Real conditions are a mobile network and a cheap laptop microphone.
Questions people ask
What is a good time to first audio for a voice agent?
Sub-second is the general design target. Past roughly 1,500 milliseconds a conversation is widely reported to feel broken. There is a floor as well as a ceiling: under about 200 milliseconds the reply lands as an interruption instead of an answer, because nobody could have considered the question that quickly. Aim for natural rather than minimal.
What is the difference between TTFA and TTFT?
Time to first token measures the language model alone. Time to first audio measures the whole system, from the moment a person stops speaking to the moment sound comes back, so TTFT is one component of it. Presenting TTFT as TTFA is the category-defining overclaim, and it matters because a full pipeline always sits above the model figure on its own.
Why is my voice agent slow when the model is fast?
Because the model is often not the largest segment of the budget. Endpointing, transport, transcription, retrieval, speech synthesis and playback all consume time before anyone hears a sound. In one familiar failure the model accounts for 700 milliseconds while retrieval accounts for 800, because an unindexed vector search is scanning a corpus that has grown since launch. Instrument every stage separately before blaming the model.
How should time to first audio be measured?
End to end, in wall-clock time, from the moment the person stops speaking to the moment the first sound of the reply plays. Quote a percentile rather than a median, because the bad calls decide how a voice product gets judged, not the average one. Measure over the networks and microphones your users actually have, not on a wired lab connection. A specification that does not say where the clock starts and stops produces vendor bids that cannot be compared.
Last reviewed 30 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.