Silence threshold
In one sentence
The silence threshold is the amount of trailing quiet a voice agent waits for before deciding the speaker has finished, usually set between 150 and 300 milliseconds, and it is the single most-felt setting in the pipeline because it is paid on every turn.
Definition
The silence threshold is how much quiet a voice agent listens for before it concludes you have stopped and it can reply.
It is the specific dial behind endpointing, and it is the number a visitor feels most sharply, because a slice of it is spent ahead of every single answer.
The threshold governs one quantity: how many milliseconds of detected quiet have to elapse before the agent calls the turn over. Conversational deployments usually land it somewhere in the 150 to 300 band. Because that wait recurs on every turn, it is among the biggest slices of time to first audio that a team can actually move.
Two failure modes, with opposite causes
- Set it too short and the agent jumps in on an ordinary pause. That hurts most when someone is reading out digits, an address or a list, all of which are full of built-in pauses.
- Set it too long and every reply drags behind an avoidable wait, which reads as a slow system even when the rest of the stack is quick.
- No single value serves both cases. That is a built-in property of judging turns by silence, not a tuning problem waiting to be solved.
Why per-prompt thresholds are the real answer
- After a yes-or-no question a short value is safe, because the reply is brief and self-contained.
- A prompt asking for a phone number wants a much longer one, because people read digits in clusters with pauses between them.
- The agents that come across as rude are usually running one global value across every context. Matching the wait to the answer you asked for is a small change that lands hard.
Signals that can shift it on the fly
- A trailing conjunction or filler word in the partial transcript, which hints the speaker has more coming.
- Pitch that holds level instead of falling, which tends to mark an unfinished thought.
- How likely it is the sentence is actually complete, which is what semantic turn detection reads and the fuller solution.
How the network gets involved
- Jitter can punch gaps into the audio that the system mistakes for silence.
- A value dialed in on a clean office link can behave differently in the wild, surfacing specifically as the agent cutting off mobile users.
Common misconception
That a shorter threshold is simply an optimization. It is one half of a tradeoff, not a free win. The right value turns on what the speaker is likely to be doing at that instant, which is exactly why a single global number is the wrong instrument.
Why it matters commercially
Getting talked over by a machine mid-sentence is a genuinely grating thing to endure, and it signals that the system is not paying attention. The fix costs almost nothing, because it is a configuration change rather than a model one.
In voice specifically
Type into a chat box and you can stall for a minute mid-answer with no penalty. A microphone gives the agent no way to tell thinking from finishing, so it falls back on the length of the silence, and it spends a fraction of this threshold making that guess before every reply.
Where AsqVox fits
AsqVox runs as a website voice agent, where a visitor speaks to navigate and get answers, so turn-taking quality is felt directly as whether the agent is really listening. Fitting the pause to the prompt is a small piece of setup with an outsized effect on how attentive the agent seems.
Visual
The number paid on every turn
No value satisfies both failure modes. That is the point.
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.
Silence 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. Shortening it buys latency and starts cutting people off mid-sentence.
The average gap between turns in human conversation is around 200 milliseconds, which is why thresholds cluster near that figure.
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 consumes roughly a third of a sub-second time to first audio target, and it consumes 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. The silence threshold is one of the easiest contributors to that total to reduce, and one of the riskiest.
Silero VAD scores a 30 millisecond audio chunk in under 1 millisecond, so the detection feeding the threshold is not itself a latency contributor.
Under 1 ms per 30 ms chunkvendor claimSilero VAD published benchmark, 2026 - Published by the project itself. The useful conclusion is directional: the wait is the threshold you set, not the time taken to notice the silence.
No published research establishes optimal silence thresholds by task type, despite that being an obviously useful and measurable question. Current practice is empirical and decided deployment by deployment.
-no reliable figureSo a vendor default is a guess about someone else's traffic, not yours. The only figures worth trusting come from measuring interruption and completion on your own prompts, one at a time.
Examples
In practice
A booking agent asks for a mobile number, and because the caller breathes between digit groups, endpointing fires on the first group. The agent reads back half a number and the exchange loops. Lifting the threshold to around 800 milliseconds for that one prompt, then dropping it back, clears the failure without slowing anything else, and task completion on the number step improves at once.
The everyday version
Picture the silence threshold as the beat the agent leaves after a customer stops before it assumes they are done. Set it too tight and it barges in partway through an address or a card number. Set it too loose and a small dead pause trails every reply. The way out is not one setting but several, chosen by whatever you just asked.
Usage
Who says it
- Voice engineers, routinely. It is among the first parameters they reach for on a new deployment.
- Non-technical operators increasingly, because it now shows up as a knob in vendor configuration panels.
Where it turns up
- In agent configuration panels, tuning guides and post-launch quality reviews.
- In those quality reviews it is usually the reason behind a complaint that the agent kept interrupting.
Common misuse
- Running one global value across all prompts. Most agents that feel rude are doing exactly this.
- Reading a lower value as an optimization instead of a tradeoff.
- Tuning on a clean office network, then shipping to mobile users where jitter changes the behavior.
Questions people ask
What is a good silence threshold for a voice agent?
Conversational systems usually 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 it comes down to the prompt: a short value after a yes-or-no question, and as much as 800 milliseconds while someone reads out a phone number.
Why does my voice agent interrupt me mid-sentence?
Most often the silence threshold is set too short, so it triggers during a natural pause, which is worst while you recite a number, an address or a list. On mobile it can also be network jitter opening gaps that read as silence. The fix is a per-prompt threshold, and for mobile, checking the network rather than only the tuning.
What is the difference between a silence threshold and endpointing?
The silence threshold is the specific setting, how much quiet the agent waits for. Endpointing is the whole decision process that uses it to declare the turn finished. The threshold is the number you tune; endpointing is the mechanism that number feeds.
Is a shorter silence threshold always better?
No. The delay is added to every turn, so shortening it does buy real latency, and a 300 millisecond threshold eats about a third of a one-second budget before any processing begins. But short thresholds cut people off exactly when it matters most, while they are carefully reading something out. It is a tradeoff, not an improvement.
Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.