Home › Topics › Jitter & Real-Time Traffic › Jitter Targets

Jitter Targets by Workload: VoIP, Video, Remote Desktop


Sooner or later you need a number. The alert-threshold field is blinking at you, a vendor is asking what jitter their service should expect on your network, or a manager wants to know whether 22 ms is "bad." And the honest answer — "it depends on the workload" — is true but useless on its own, because nobody hands you the dependency table. This guide is that table, with something more valuable attached: the reasoning. That lets you defend each number and adjust it intelligently instead of cargo-culting it.


The core insight is that jitter tolerance is not a property of networks. It is a property of applications, and specifically of how much buffering each application can afford. A voice call, a video meeting, a remote desktop session, and a streaming movie can share one wire. They can experience the same 40 ms of jitter as, respectively, a disaster, a nuisance, an annoyance, and nothing at all. Same network, four verdicts. That is why a single "good jitter" number does not exist and why anyone who quotes one without asking what you are running is guessing.


One prerequisite: these targets assume you are measuring honestly — the smoothed-average style figure most tools report, sanity-checked against the spikes users actually feel. If your measurement habits are shaky, detour through measuring jitter properly first. A solid threshold applied to a meaningless number alerts on noise and sleeps through outages.


In this guide

  • Where these numbers come from
  • The master table
  • Voice calls: the most sensitive
  • Video conferencing: a little more forgiving
  • Remote desktop and interactive sessions
  • Streaming: why Netflix works when calls don't
  • Turning bands into alerts
  • The caveats that keep you honest

Where these numbers come from


Let us be upfront about the epistemology, because it will make you better at using the table. Jitter targets are not laws of physics. Nobody derived "30 ms" from first principles and carved it anywhere. What the bands actually are is engineering folklore made explicit: the compressed experience of a great many operators. They rest on three legs of real arithmetic. First is the packet cadence of the workload. Voice sends every 20 ms, so jitter comparable to the cadence means packets arriving into the wrong time slots entirely. Second come the buffering norms of the receiving apps. A jitter buffer that typically runs tens of milliseconds deep can absorb jitter smaller than itself. It must convert anything larger into delay or discards — the machinery from the jitter buffers guide. Third is the human delay budget: roughly 150 ms one-way before conversation degrades. That caps how deep buffers may grow, which caps how much jitter is absorbable.


Cadence, buffer norms, delay budget: every band below is those three quantities colliding for one workload. That is what makes the numbers defensible. It also tells you when to deviate from them. If one leg changes on your network, the band moves with it and you will know which way. The change might be an unusual codec cadence, an app with atypical buffering, or a path with no delay headroom.


The master table


Values are end-to-end jitter on the path the workload actually uses, from the user's segment to the far end. "Excellent" means the workload never notices the network. "Acceptable" means it works, with the coping machinery visibly earning its pay. "Investigate" means users are feeling it, whether or not they have filed the ticket yet.


WorkloadExcellentAcceptableInvestigate
Voice calls (VoIP)< 10 ms10–30 ms> 30 ms sustained
Video conferencing< 15 ms15–40 ms> 40 ms sustained
Remote desktop / interactive< 20 ms20–50 ms> 50 ms sustained
Streaming / one-way video< 50 ms50–200 msonly if rebuffering appears
Fast-paced online gaming< 10 ms10–30 ms> 30 ms sustained

Two notes on applying it. The numbers refer to the everyday jitter figure your tools report — the smoothed or averaged style. So pair them with a glance at the p95 and maximum from the same window. A path whose average sits politely in "acceptable" while its p95 lives in "investigate" belongs in "investigate." That is because users experience the spikes, not the average. And measure during representative hours; a band verdict earned at 7 AM on an idle link certifies nothing about the 10 AM all-hands call.


Now the part that makes the table trustworthy: why each row sits where it sits.


Voice calls: the most sensitive


Voice earns the strictest row because all three legs of the arithmetic are unforgiving at once. The cadence is a packet every 20 ms, so 30 ms of jitter means arrival timing is wandering by more than a whole packet slot. The stream's rhythm is not merely smudged but scrambled. The buffers in phones and softphones typically operate in the 20–80 ms range. So jitter up to about 30 ms lets an adaptive buffer sit comfortably at 40–60 ms of depth and absorb everything. And the delay budget is tight. Typical network delay and codec processing already spend most of the ~150 ms conversational allowance. There is simply no room to buffer away much more jitter. Push jitter past the band and the buffer must pick a failure: discards (robotic audio) or depth (talk-over). There is no third option — that is the trade-off triangle doing exactly what it must.


Hence the bands. Under 10 ms, buffers idle at shallow depths and voice is indistinguishable from a landline. This is what healthy wired LANs and good WAN paths deliver routinely, so treat it as the standard, not a stretch goal. From 10 to 30, calls work and users rarely comment, but the margin is spent. Any additional insult — a loss burst, a congestion spike — now lands on a system with no reserves. Past 30 sustained, complaints are a matter of time. Voice also travels with two companion thresholds worth pinning next to it. Keep one-way delay inside roughly 150 ms and packet loss under about 1%. The reasoning for that one lives in the acceptable packet loss guide. That is because the three budgets are spent from the same purse. A path at 140 ms of delay cannot afford deep buffering, so its effective jitter tolerance shrinks below the table's number.


Video conferencing: a little more forgiving


A video meeting is really two streams with different constitutions sharing a session. There is an audio stream that is exactly the voice case above, and a video stream that is bulkier but better at hiding damage. Video tolerates more for two reasons. First, its material is more concealable. When packets are late, the app can hold the last frame a beat longer, drop a frame, or quietly step the resolution down. Viewers forgive all of these far more readily than they forgive mangled speech. Second, conferencing apps run somewhat deeper and more elastic buffering on the video path. That is because a few extra tens of milliseconds on pictures barely dents the conversation so long as the audio stays tight.


That is also why the failure sequence past the band is so recognizable. Mild excess: motion gets subtly choppy and resolution breathes up and down as the app trades sharpness for survival. Worse: frames freeze outright, then the picture leaps to catch up. And notably, audio usually breaks before video becomes unwatchable. That is partly because your ear is the harsher judge, partly because the audio stream has the least buffer room to hide in. Users report it in exactly those terms: "the video was frozen but we kept talking" is a mildly stressed network. "The sound went robotic" is a badly stressed one. The band — excellent under 15, acceptable to 40 — is the voice band widened by video's concealment tricks. It is worth remembering that when a meeting degrades, the audio half of it is still living on the voice row.


Remote desktop and interactive sessions


Remote desktop protocols are a different animal. They are not a fixed-cadence media stream but a conversational exchange of small events — keystrokes and clicks up, screen updates back. There is no strict 20 ms metronome and no concealment cliff, which buys the row extra width. What jitter attacks instead is the feel. Human motor control adapts astonishingly well to a consistent delay: people settle into a steady 60 ms echo within minutes and stop noticing it. What nobody adapts to is inconsistency. In a session where the echo is 20 ms for one burst and 90 for the next, typing feels broken and "laggy." That is true even when the average is modest, because your fingers can never find the rhythm. Jitter here is experienced as unpredictability of touch: cursor rubber-banding, scroll stutter, that maddening double-keystroke when you re-press a key you believed had not registered.


The arithmetic legs land accordingly. Interactive protocols buffer very little by design. Buffering keystroke echo defeats the purpose of the session. So jitter passes through to the user's hands more directly than through a media buffer. But with no audio-style cliff, degradation is gradual rather than catastrophic. Under 20 ms of jitter, sessions feel local-ish and nobody thinks about the network. From 20 to 50, fine work — precise cursor placement, fast typing, drawing — starts to feel faintly drunk while reading and clicking remain fine. Past 50 sustained, users describe the session as unusable for real work well before file transfers or video would care. If your organization lives in virtual desktops, this row, not the voice row, is your daily quality bar.


Streaming: why Netflix works when calls don't


The streaming row looks like a typo — hundreds of milliseconds of jitter, shrugged off. It is the most instructive row in the table, because the households and branch offices you support see the contrast weekly. Movie streaming is flawless all evening, yet the morning's video call was a slideshow. Everyone concludes the call app must be broken. The network was likely the same. The difference is the delay budget, and it changes everything.


A one-way stream is television, not conversation. Nobody on the sofa cares that the frame they are watching left the server thirty seconds ago. There is no turn-taking to protect. So the player buffers seconds to tens of seconds of video. That is a shock absorber a thousand times deeper than a call's, refilled opportunistically in bursts over TCP. Jitter of 200 ms is invisible against a thirty-second reservoir; even a multi-second stall just drains the buffer slightly while playback rolls on serenely. Streaming only visibly fails — rebuffering wheel, quality collapse — when average throughput can no longer keep the reservoir filled. That is a bandwidth problem, not a jitter problem. That is the whole lesson in one line: the moment you must react in real time, the buffer must be shallow, and every millisecond of jitter matters; remove the interactivity and jitter almost stops existing. When a user says "streaming works fine, so the network is fine," you now have the two-sentence answer: streaming is allowed to run half a minute behind reality; your call is allowed about a tenth of a second.


The gaming aside: fast-paced multiplayer games sit at the strict end for the same structural reason as voice. They exchange small, frequent state updates that are only useful for an instant, so nothing meaningful can be buffered. Jitter shows up as rubber-banding and hit-registration weirdness. Gamers will report a link that "pings fine" but plays badly — often the household's most sensitive jitter detector. Treat competitive gaming like voice: excellent under 10, trouble past 30.


Turning bands into alerts


A threshold that lives in a guide helps nobody at 2 PM on a Tuesday. It needs to live in your monitoring system with an alert attached. The clean pattern is to make the alert levels mirror the bands: warn where "acceptable" ends, critical comfortably past it. Always use a sustain requirement, because single-poll jitter spikes are weather, not climate:


Path carriesWarn (sustained 10–15 min)Critical (sustained)
Voice / softphones> 30 ms> 50 ms
Video conferencing> 40 ms> 60 ms
Remote desktop / VDI> 50 ms> 80 ms
Any path, regardless of workload> 3× its own baseline> 5× baseline, or still climbing

Three habits make these fire usefully. First, the sustain window: alert when the figure stays over the line across consecutive polls, not on one bad sample. You want to be told about the storm, not every raindrop. Second, the baseline row is not decoration. A path that lives at 2 ms and jumps to 12 is telling you something changed even though 12 is "acceptable" on paper. Catching that drift early is how you fix causes before users feel effects. Third, measure where the workload lives. An end-to-end check along the real path (office to phone provider, branch to VDI farm) mirrors user experience. It is the number these bands apply to. Per-segment checks (LAN hop, WAN hop, Wi-Fi hop) are for assigning blame after the end-to-end alarm fires. Each segment should be consuming only a fraction of the band on its own.


If you are rolling this out from scratch, the sequence is short:


  • 1. Inventory the real-time paths. List where voice, conferencing, and remote desktop traffic actually flows — office to provider, branch to VDI farm, site to site. Most networks have three to six paths that matter.
  • 2. Measure each one properly. A representative window, both directions, median/p95/max — the methods are in measuring jitter. Write the healthy values down; that is your baseline.
  • 3. Set warn and critical per path using the strictest workload the path carries — a link that carries any voice gets the voice thresholds.
  • 4. Let it soak for two weeks, then tune. A warn that fired daily without a matching complaint is set too tight for your normal. Raise it or fix the path, but do not learn to ignore it.

Tip: alert thresholds are also communication tools. When you set warn at 30 ms on the voice path, write the reasoning — cadence, buffer norms, delay budget — into the alert's description field. Six months from now, whoever sees the alert (including you) will know what it means and whether the number deserves updating. They will not treat it as inherited scripture.


The caveats that keep you honest


Bands are end-to-end; segments must share them. A call's jitter budget is spent along the whole path: the user's Wi-Fi, the LAN, your WAN edge, the provider. Two segments that are each "acceptable" on their own can add up to a path that is not — variability compounds, even if not by simple addition. Engineer each segment to sit well inside the band. Reserve the largest share for the segment you least control, which is usually the last-hop wireless. The Wi-Fi packet loss guide explains why radio will always be your wobbliest link.


Your normal beats generic normal. These bands are defensible defaults for judging a path cold. But a measured baseline of your paths — median, p95, max per hour of day — outranks any table. That is because "different from our normal" catches developing problems that "worse than folklore" misses by a mile. Use the table to set the guardrails; use your baseline to notice the drift.


Folklore, made explicit, is still folklore. If a vendor contract, an app's published requirements, or your own testing gives you a different number for a specific system — believe the specific number. What this guide gives you is the reasoning to evaluate such claims. A vendor demanding jitter under 5 ms for a chat app deserves a raised eyebrow. One asking for under 20 on a real-time trading floor deserves a nod. Defend the arithmetic, not the digits.


Where to go next


The table in one breath: voice is strictest because a tight cadence, shallow buffers, and a hard human delay budget leave nowhere to hide; video conferencing buys slack with concealment but its audio still lives by voice rules; remote desktop suffers in feel rather than in cliffs; streaming barely cares because it is allowed to run half a minute behind reality; and every band is cadence, buffers, and budget colliding — folklore you can now defend with arithmetic.


If your paths are inside the bands, set the alerts and enjoy the quiet. If they are not, diagnosing jitter is the isolation sequence for finding the source. The guide QoS for real-time traffic is how you protect calls from the traffic you cannot remove. And MOS scores shows how these numbers roll up into the quality score users would give. The rest of the Jitter & Real-Time Traffic topic ties the whole story together. If delay itself is your bigger worry, the Network Latency topic covers that budget in full.


Put this into practice with SysPeep Monitor

Bands only earn their keep when something is watching them. SysPeep Monitor auto-discovers your network. It polls response times, link utilization, and device health over SNMP and WMI. It lets you set warn and critical thresholds with sustain windows and dependency-aware alert rules. So drift from your baseline gets noticed before the meeting-room complaints start. The free trial installs on a Windows machine in minutes.

Download Free TrialLearn About SysPeep