Jitter

Latency & Performancealso: packet delay variation

In one sentence

Jitter is variation in the timing of arriving packets, and because audio playback needs a steady stream, a packet that arrives late is as damaging as one that never arrives at all, since the moment it was needed has already passed.

Latency vs dropped packets Real-time audio engineering convention, 2026Last reviewed 31 July 2026

Not to be confused with Round trip time.

Definition

Jitter is variation in the timing of arriving packets. A steady delay is manageable; a delay that keeps changing is what makes audio stutter and break up.

It bites because playing audio back needs a steady stream, so a packet that lands late is as useless as one that never lands, because its moment has already gone.

Jitter matters because audio playback needs a steady feed of packets. A late packet is as damaging as a missing one, since by the time it arrives the moment it was meant to fill has passed.

Where jitter comes from

  • Queuing that varies at network hops as load changes.
  • Routes that change partway through a session.
  • Wireless interference and signal swings, which make it much worse on mobile and on congested Wi-Fi.
  • Other traffic competing for a shared connection.

The jitter buffer, and the trade it forces

  • Incoming packets are held for a short spell before playback, so late arrivals get a chance to catch up and play in order.
  • A bigger buffer soaks up more jitter and adds that same delay to every packet, including the ones that were already on time.
  • A smaller buffer keeps latency low and drops more late packets, which leaves audible gaps.
  • Adaptive jitter buffers move with the conditions they observe, and that is standard practice in modern real-time audio.

Why voice agents feel it more than most

  • Every millisecond of jitter buffer is added to a conversational latency budget that is already tight.
  • A buffer sized generously for bad networks hurts the experience for everyone on good ones.
  • Jitter hits the visitor's audio reaching the system and the agent's audio reaching the visitor, and the two directions can differ.

The tie-in with voice activity detection

  • Jittered audio can open gaps that voice activity detection reads as silence.
  • That can trip premature endpointing, so a network fault shows up looking like the agent cutting the visitor off.
  • Diagnosing a turn-taking complaint therefore means checking the network, not only the endpointing settings.

Common misconception

That jitter is just another name for latency. A steady 200 millisecond delay is workable. A delay swinging between 50 and 400 milliseconds is not, even though its average comes out lower.

Why it matters commercially

Jitter decides whether voice holds up on the networks visitors actually use, mobile most of all. It is invisible in an office test on a wired connection, which is exactly where it stays hidden until launch.

In voice specifically

A file download rides out jitter without anyone noticing, because it can wait for the stream to reassemble in order. A live conversation has no such luxury: the audio has to play the instant it is due, so a packet that arrives out of rhythm either leaves a gap or gets thrown away.

Where AsqVox fits

AsqVox runs in the browser, so it inherits whatever network the visitor is on with no control over it. Buffering that adapts to conditions, and a text chat fallback for when audio cannot hold up, are the responses available to it.

Visual

Late is as bad as lost

Late is as bad as lostHow packets arriveWhat playback doesWhat the listener getsConsistent delayEvery packet late by the same amount, ateven intervalsRuns smooth and continuousA steady 200 ms delay. Workable.JitteredArrivals scatter, some early and somevery lateBreaks wherever a late packet missed itsmomentA lower average delay, a worseexperience.

Steady delay is a cost. Variable delay is a defect.

Both streams carry the same packets; only the timing differs. The jitter buffer sits between them and the trade it makes is unavoidable: a small buffer keeps latency low and drops late packets into audible gaps, a large buffer plays smoothly and adds its full depth to every packet, and every millisecond of it lands on a latency budget that is already tight. Adaptive buffers move with the network, which is standard now. The misdiagnosis to watch for: a jitter gap reads to voice activity detection as silence, which trips premature endpointing, so a network fault presents as the agent interrupting the visitor.

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.

Jitter buffers trade added latency against dropped late packets, and adaptive buffers adjust to observed conditions dynamically, which is standard practice in modern real-time audio.

Latency vs dropped packetsindustry range

Real-time audio engineering convention, 2026 - A mechanism fact rather than a measured figure. It tells you the trade exists, not where any given deployment has set its buffer.

WebRTC includes congestion control and packet loss concealment designed for speech, which handle jittered and lossy conditions better than plain transport.

Built-in concealmentindustry range

WebRTC transport design, 2026 - An architectural fact about how real-time media moves, not a quality number. It is why browser voice weathers a bad network more gracefully than a plain connection would.

Sub-second time to first audio is the working design target for conversational voice, and jitter buffer depth is added directly to that budget.

Under 1,000 msindustry range

Industry working threshold, 2026 - A working convention rather than a published standard. The point for jitter is that every millisecond spent smoothing the stream is spent against this target.

Past roughly 1,500 milliseconds, a conversation is widely reported to feel broken.

1,500 msindustry range

Industry working threshold, 2026 - The same shared working line the rest of the latency category uses. A generously sized jitter buffer can push a borderline connection over it.

Silero VAD scores a 30 millisecond audio chunk in under 1 millisecond, so voice activity detection is fast enough to react to a jitter-induced gap, which is precisely why jitter can trigger false endpointing.

Under 1 ms per 30 ms chunkvendor claim

Silero VAD published benchmark, 2026 - Published by the project itself. The directional point is that detection is quick enough to mistake a network gap for the end of a turn.

No published benchmark measures voice agent performance under controlled jitter conditions across platforms.

-no reliable figure

So the effect of jitter on a real deployment is something you can only observe on your own stack, over the networks your visitors actually use.

Acceptable jitter thresholds for conversational voice are not standardized in published form, and they vary with the buffer strategy in use.

-no reliable figure

A spec that promises to tolerate a given jitter figure is promising against a line no standards body draws, and the buffer choice behind it changes what that figure means.

Examples

In practice

Mobile visitors report the agent cutting them off mid-sentence. The endpointing configuration checks out, and the same visitors have no trouble on Wi-Fi. Network analysis shows jitter opening gaps in the audio stream that voice activity detection reads as end-of-speech. Raising the minimum speech-duration requirement and adjusting how the buffer adapts clears it. The turn-taking configuration was never the problem.

The everyday version

Jitter is audio arriving unevenly instead of in a steady stream. It is why a call can sound choppy even when the connection looks fast, it is far worse on mobile, and it does not show up at all when you test the system on the office broadband.

Usage

Who says it

  • Network and real-time audio engineers, precisely.
  • Contact center operations, in call-quality diagnostics.
  • Rarely a business term, though it surfaces as complaints about choppy audio.

Where it turns up

  • In a spec sheet, next to network requirements, jitter buffer behavior, packet loss tolerance and the supported network conditions.

Common misuse

  • Treating jitter as the same thing as latency.
  • Diagnosing turn-taking complaints without ever checking the network conditions behind them.
  • Testing only on wired connections, where jitter barely appears.

Questions people ask

What is the difference between jitter and latency?

Latency is delay; jitter is variation in that delay. A steady 200 millisecond delay is workable, because playback can settle into a rhythm around it. A delay that swings between 50 and 400 milliseconds is not workable, even when its average comes out lower, because audio playback needs a steady stream and a packet that arrives out of rhythm either leaves a gap or gets discarded.

Why is my voice agent choppy only on mobile?

Because cellular and congested Wi-Fi connections produce far more jitter than a wired office network, and jitter is what makes audio stutter and break up. The office test on broadband will look clean and hide the problem entirely. Check the jitter buffer behavior and the network conditions your actual visitors are on, not only the ones you tested on.

What is a jitter buffer?

A jitter buffer holds incoming packets for a short spell before playing them, so late arrivals have a chance to catch up and play back in order. It forces a trade: a larger buffer absorbs more jitter and adds that delay to every packet, including the ones that were on time, while a smaller buffer keeps latency low and drops more late packets into audible gaps. Adaptive jitter buffers adjust to the network as it changes, which is standard in modern real-time audio.

Can jitter make a voice agent interrupt people?

Yes, and it is a common misdiagnosis. Jitter can open gaps in the audio stream that voice activity detection reads as silence, which trips premature endpointing, so the agent appears to cut the visitor off mid-sentence. The cause is the network rather than the endpointing setting, which is why a turn-taking complaint should be checked against network conditions before the silence threshold is retuned.

Share this definition

Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.