Endpointing
In one sentence
Endpointing is how a voice agent decides that the person has finished speaking and its own turn has begun, normally by timing the silence against a threshold set between 150 and 300 milliseconds.
Not to be confused with Voice activity detection, Turn detection, or Barge-in.
Definition
Endpointing is how a voice agent works out that you have finished talking and it is now allowed to answer.
It is normally just a timer. Go quiet for a set number of milliseconds and the turn is closed. No other setting in a voice agent is felt as directly as this one.
The rule sits on top of the voice activity detection signal and is close to the simplest thing that could work: if speech has been absent for N milliseconds, close the turn, finalize the transcript and start processing. For conversational systems N is usually set somewhere between 150 and 300 milliseconds.
Why the tradeoff is real and cannot be tuned away
- A short threshold is fast and responsive and reads as alert. It also cuts off anyone who pauses to think, to breathe, or in the middle of a list. Anybody reciting a number, an address or a card detail pauses naturally, and gets interrupted for it.
- A long threshold is patient and lets people get to the end of a sentence. It also drags, because that wait is now attached to every turn before anything at all begins.
- No value satisfies both. This is not an engineering failure waiting for better tuning, it is a structural limit of deciding turns by silence, and it is the reason semantic turn detection exists at all.
What it costs in the latency budget
- The wait is additive latency, paid on every turn. At a 300 millisecond threshold, no reply can arrive less than 300 milliseconds above the theoretical floor.
- Of everything inside time to first audio, this is frequently the largest segment you can actually control, so it is the first thing to check when a system feels slow.
Adaptive and context-aware approaches
- Vary the threshold by the kind of answer you just asked for. Short after a yes-or-no question, long after asking for a phone number.
- Read the partial transcript before deciding. A trailing "um", or a sentence that ends on "and", suggests more is coming.
- Both are pragmatic improvements to a silence-based approach. Neither is semantic turn detection, which is a different mechanism.
Common misconception
That you are aiming for the smallest number you can survive. What you are actually aiming for is a number that fits whatever the visitor is likely to be doing right then. Tune everything for quick back-and-forth and the system will fail reliably at the moments that carry the most value, which are the ones where somebody is slowly reading a number out.
Why it matters commercially
When somebody calls a voice agent rude, this setting is usually the reason. Having a machine talk across you halfway through a sentence is a particularly unpleasant thing to sit through, and what it communicates is that the system does not care. Correcting it costs close to nothing, which makes it one of the cheapest quality wins available.
In voice specifically
A chat box will wait as long as you like while you compose. Speech gives the system no equivalent signal that you are still thinking, so it has to infer the end of your turn from silence, and it pays for that inference on every turn.
Visual
The setting you cannot get right for everyone
300 milliseconds of politeness, charged on every turn.
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.
Endpointing thresholds in conversational voice systems typically sit between 150 and 300 milliseconds.
150 to 300 msindustry rangeCommon production budgeting convention, 2026 - An engineering convention drawn from production practice, not a published research finding, and it should be quoted that way.
The average gap between turns in human conversation is around 200 milliseconds, which is why thresholds cluster near that figure. A machine that waits noticeably longer than a person would is perceptibly slower than a person.
200 msindustry rangeWidely cited conversational turn-taking baseline, 2026 - A reference point rather than a target. It explains where the conventional band came from, and it is not a measurement of any system you might buy.
A 300 millisecond threshold takes about a third of a sub-second time to first audio target, and it takes it before a single processing step has run.
About one third of the budgetindustry rangeIndustry working threshold, 2026 - The sub-second target is itself a working convention rather than a published standard. The share is arithmetic on top of it, so treat it as a planning figure.
Past roughly 1,500 milliseconds in total, a conversation is widely described as feeling broken.
1,500 msindustry rangeIndustry working threshold, 2026 - A shared working line rather than a measured cliff. Endpointing is one of the easiest contributors to that total to reduce, and one of the riskiest.
Nobody has published research identifying the best threshold for a given task type, which is an obviously useful thing to know and still nobody has it. Practice today is empirical and decided deployment by deployment.
-no reliable figureSo a vendor default is somebody else's guess about your traffic. The only way to find your own numbers is to measure interruption and completion on the prompts that matter, one prompt at a time.
Examples
In practice
A booking agent asks for a mobile number. Callers read the digits out in groups with a natural breath between them, and the 250 millisecond threshold fires after the first group. The agent answers a half-finished number, the caller corrects it, and the exchange goes round again. The fix is to lift the threshold to around 800 milliseconds for that one prompt and restore it afterwards. Completion on the number-capture step improves immediately.
The everyday version
Endpointing is the pause the agent allows after you stop talking before it concludes you have finished. Make that pause too brief and it will talk across you halfway through your address. Make it too generous and every sentence you say is followed by an uncomfortable gap.
Usage
Who says it
- Voice engineers, constantly. It is one of the first things they reach for when tuning a new deployment.
- Buyers now meet the word too, because it increasingly appears as a setting in vendor configuration panels aimed at non-technical users.
Where it turns up
- In tuning guides, integration documentation and the agent configuration panel itself.
- In post-launch quality reviews, where it is usually the answer to why people say they were interrupted.
Common misuse
- Leaving it as one global number rather than setting it per prompt. Most systems that come across as rude are running a single threshold across every context.
- Confusing it with barge-in. Endpointing decides when the visitor's turn ends. Barge-in handles the visitor starting to speak during the agent's turn.
- Assuming a lower value is an optimization. It is a tradeoff, and moving it buys one thing by spending another.
Questions people ask
What is a good endpointing threshold?
Conversational systems typically sit between 150 and 300 milliseconds, which is an engineering convention rather than a research finding. There is no published optimum by task type, so the honest answer is that it depends on the prompt: short after a yes-or-no question, and as much as 800 milliseconds while somebody reads out a phone number.
Why does my voice agent interrupt me when I read out a phone number?
Because people recite digits in groups with a natural pause between them, and a 250 millisecond timer fires in that pause. The agent then answers a half-finished number. The fix is not a better model, it is raising the threshold for that one prompt and putting it back afterwards.
What is the difference between endpointing and barge-in?
Endpointing decides when the visitor's turn has ended so the agent can start. Barge-in handles the opposite direction: the visitor starting to speak while the agent is already talking, and the agent stopping for them. They are frequently used as if they were the same thing, and they solve different halves of turn-taking.
Is a shorter endpointing threshold always better?
No. The delay is additive on every single turn, so shortening it does buy real latency, and a 300 millisecond threshold eats roughly a third of a one-second budget before any processing begins. But short thresholds cut people off exactly when it matters most, which is while they are carefully reading something out. It is a tradeoff, not an improvement.
Last reviewed 30 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.