DeviceBenchFree network jitter test
Jitter test for the wobble an average ping hides
A call breaks up on a line whose ping test reads a steady 18 ms, and neither reading is wrong: what a listener hears is not the average delay but how much it changes from one packet to the next. Sixty round trips 150 ms apart are measured here and each is drawn as a bar showing how far it landed from the one before it, with the reading given as the mean change between neighbors rather than a standard deviation — a line alternating 20 ms and 80 ms scores the same as one drifting gently between them, and only the first ruins a conversation. The probes are web requests to this domain rather than ICMP, the same measurement the ping test explains in full.
- 100% free
- No signup
- 60 probes
- Mean absolute change
- Buffer depth estimated
Nine seconds of probes, 150 ms apart. Leave the tab in front — a hidden tab has its timers slowed down and the spacing stops being what it says.
Press the button and the run takes nine seconds. The bands are 5 ms, 20 ms and 50 ms of average change.
After a run this converts the spread into the delay a receiver has to add to hide it, counted in 20 ms audio frames.
Run it twice: once with the house quiet, once while a phone is backing up or someone starts a stream. The pair of numbers tells you whether the wobble belongs to your line or to whatever else is sharing it.
How to measure and chase down network jitter
Nine seconds per run, so the method is to change one variable at a time and keep running it.
Measure from where you actually sit
Take the reading at the desk, the sofa or the kitchen table you make calls from, not next to the access point. Wireless variation is a question about the path between one radio and another, so a laptop two rooms and one brick wall away from the router produces a different chart from the same laptop at arm's length — and the room you use is the one that matters.
Read the bars, not only the headline
A dense hedge of small bars is a healthy line: consecutive probes are landing close together and a receiver has almost nothing to smooth out. Three or four tall amber spikes in nine seconds are the shape that ruins speech, even when the average is respectable, because a listener notices the gap and never notices the mean. The tallest single step is printed beside the average for that reason.
Change exactly one thing, then run it again
Move from 2.4 GHz to 5 GHz. Plug in a cable. Pause the backup. Unplug the powerline adapter. Each run takes nine seconds, so testing four variables properly costs under a minute, and a single change between runs is what makes the comparison mean anything. Write down the number after each — the one that halves it is your culprit.
Technical specifications
| Probes per run | 60, spaced 150 ms apart — nine seconds of continuous sampling, which is long enough to catch a wireless retry storm |
|---|---|
| Quantity reported | Mean absolute difference between consecutive round trips: packet delay variation as RFC 3550 defines it, without the exponential smoothing that specification then applies |
| Bands | Under 5 ms even, 5-20 ms a slight wobble, 20-50 ms rough, above 50 ms broken up |
| Shown alongside | Largest single step, median latency, and the standard deviation — kept off the headline on purpose, because it answers a different question |
| Buffer estimate | 95th percentile minus the fastest probe, converted into 20 ms audio frames; a round trip carries both directions, so treat it as a ceiling |
| Unanswered probes | Left out of the series rather than counted as a change — a gap is not a delay variation, and averaging across one would invent a spike |
| Data kept | None. Sixty timings exist for the length of a run and are gone on reload |
| Price | Free, no signup, and a full run costs nine seconds |
Frequently asked questions
What is a good jitter figure?
Under 30 ms is the figure networks are engineered to for carrying voice, and under 5 ms is what a wired connection to a nearby edge normally gives you. Between 5 and 20 ms is ordinary Wi-Fi and no one will hear it. Past 30 ms a receiver has to hold audio back long enough that the delay itself starts to be noticeable, and past 50 ms it stops being able to cover the gaps at all, which is when speech begins arriving in pieces.
Why is this not just the standard deviation of my ping?
Because the two statistics disagree in exactly the case that matters. Imagine two nine-second runs: one drifts smoothly from 20 ms up to 80 ms as a download ramps, the other alternates 20, 80, 20, 80 from probe to probe. Their standard deviations are identical. Only the second one makes a conversation unlistenable, because what a receiver has to absorb is the change between neighboring arrivals, not their scatter around a mean. This page reports the mean absolute difference between consecutive probes and prints the standard deviation next to it so you can watch them disagree on your own line.
Why does Wi-Fi have so much more of it than a cable?
Because a radio channel is one conversation shared by everybody on it. Before transmitting, a device listens, and if the channel is busy it waits a randomly chosen interval and listens again; a frame that is not acknowledged is retried at the link layer. All of that is invisible to your operating system and arrives as milliseconds that vary from packet to packet. Add a neighbor's access point on an overlapping channel — in the 2.4 GHz band only channels 1, 6 and 11 avoid overlapping at all — and the waiting grows without anything in your house changing.
My spikes arrive on a regular rhythm, roughly once a minute.
Regularity points at a scheduled job rather than at congestion. The usual suspects are a laptop scanning for other wireless networks, which takes the radio off your channel for a tenth of a second at a time, a backup or sync client waking on a timer, a smart TV or console checking for updates, and a DHCP lease renewal. Run the test while watching the chart and note the interval between spikes: matching that interval to something on your network is normally faster than changing hardware.
Does a jitter buffer not fix all of this?
It hides it, and charges you delay for the service. A receiver holds arriving audio back by roughly the spread it has been seeing, so that frames can be played out at a steady rate even when they arrive unevenly, and modern implementations grow and shrink that buffer continuously. The cost is that everything you say reaches the other person later. When the buffer cannot grow far enough, late frames are discarded instead, and a discarded frame sounds exactly like a lost one — which is why the panel converts your measured spread into the number of 20 ms audio frames a receiver would have to hold.
Do powerline adapters make it worse?
Frequently, and it is the least suspected cause in the list. Powerline networking sends data over mains wiring that was never designed to carry it, sharing the pairs with everything else in the building: a motor starting, a phone charger, a dimmer switch and an induction hob all inject noise, and adapters on different circuits or opposite phases of the supply have to work much harder. A steady low reading that turns spiky at particular times of day, with no wireless involved at all, is the pattern worth suspecting them for.
Where do the sixty timings end up?
In this tab, and nowhere else. Each probe is a request for a three-byte file on this domain; the time it took is kept as a number, used to draw one bar, folded into the averages, and dropped when you leave. Nothing is stored between runs, which is why the comparison in step three asks you to write the numbers down.
About delay variation and the buffer that pays for it
Jitter is a receiver's problem before it is a network's problem. Audio and video have to be played out at a fixed rate — a frame every 20 ms, a picture every 16.7 ms — while the packets carrying them arrive whenever the network gets around to delivering them. If those arrivals are evenly spaced, playback is straightforward. If one arrives 5 ms early and the next 60 ms late, something has to smooth the difference or the output stutters. RFC 3550, the specification behind almost every real-time stream on the internet, names the quantity precisely: the difference in transit time between consecutive packets, averaged. That is deliberately a measure of change between neighbors, not of spread around a mean, and it is why this page reports what it does.
The smoothing is done by a jitter buffer, and its behavior explains most of what people experience. A receiver holds incoming frames briefly, releasing them on a steady clock, and sizes that hold to cover the variation it has been seeing; when the line gets rough the buffer grows, and every millisecond it grows is added to how long your voice takes to reach the other person. Because the whole one-way budget for a conversation that feels natural is about 150 ms, a buffer that has grown to 120 ms has spent it, and the call starts to feel like a radio interview with an overlap on every turn. Worse, a buffer has a ceiling: past it, frames arriving late are simply thrown away, and a discarded frame is indistinguishable from one that was lost on the wire. That is the point at which measuring variation stops being an academic exercise — you are watching the mechanism that will decide whether your next meeting works. The video call test checks the devices at your end of the same call in one pass.
Household causes cluster into three groups, and each has a signature on the chart. Shared wireless: everyone on a radio channel takes turns, so contention from your own devices or a neighbor's access point produces a constant background of medium bars that rises as the evening goes on. Scheduled activity: a backup, an update check or a wireless scan makes a few very tall bars separated by calm, and the interval between spikes usually identifies the culprit. Filled queues: an upload saturating your uplink makes every bar tall at once and pulls the median up with it, which is bufferbloat and belongs to your own router. A fourth, less obvious one is powerline networking, which inherits the electrical noise of the whole building. Once you know which shape you have, the ping test tells you the floor your line reaches when it is behaving, and the packet loss test puts it under load to see whether the wobble turns into outright failures.
Where the sixty timings live
Every number on this page is worked out by JavaScript running in the tab you are reading it in. Nothing you type, paste or open is uploaded, logged or kept, which is also why the tools carry on working after you disconnect from the network.
Every probe in a run is a request for the same three-byte file on this domain under a fresh URL string, and the only thing kept from it is how long it took. Those sixty numbers draw the chart, produce the averages and then vanish — there is no run history, no stored best and nothing to compare against between visits except what you copied out.