How Accurate Is an Online Decibel Meter?

Somewhere between roughly ten decibels too low and ten decibels too high. That is the honest span for a browser page reading whatever microphone you happen to own, and everything below is an argument for why it is a band rather than a figure — and why a band is still useful.

What has actually been found

Chucri Kardous and Peter Shaw, both at NIOSH, compared ten iOS sound measurement apps against a reference instrumentKardous & Shaw.

Three of ten came within ±2 dBA; seven did notKardous & Shaw. Those were native apps, on one hardware platform, written by people who could test against the actual phones in their hands.

A web page has less to work with than that. It cannot ship a per-device calibration profile, it does not know which handset it is running on, and it reads the microphone through a browser audio stack rather than the operating system's own. So this site publishes the wider band the study's spread implies, not the narrow one its best app achieved.

Where the error comes from

It is not one problem. It is four, and they fail in different directions, which is why they do not cancel.

The microphone was built for speech. Its job is to make a voice intelligible on a call, which is a different job from reporting a level. Its frequency response is not published by anyone, and it is shaped for the band speech lives in rather than flat across the range a level is computed over.

Automatic gain control fights you. A phone microphone is normally applying its own gain, turning quiet things up and loud things down — which is exactly what a meter must not do. A browser can ask for it to be switched off, and this one does, but the request is a request. When a device refuses, this meter says so on screen rather than quietly reporting a compressed number.

There is headroom, and then there is not. Phone inputs run out somewhere around 100 to 105 dB. Past that the number stops climbing while the room keeps going, and a reading that has flattened out should be read as at least that, never as exactly that. This meter warns when the input looks like it is clipping, for exactly this reason.

The curve is exact; what it is applied to is not. The A curve is evaluated band by band from its published response, so the weighting itself is no longer a source of error. What remains uncertain is everything underneath it: a spectrum computed in short windows, from a microphone whose own response across that range is unknown. A correct curve over an unknown input is still an unknown number, and the curve being right is the smallest of the four problems.

None of the four is fixable in a browser. A page that claims otherwise is claiming something about your hardware that it has no way to know.

Why a band and not a correction

The obvious fix is to guess. Detect the device, look up a typical offset for that class of hardware, and quietly subtract it. This site does not do that, and the reason is worth stating because it looks like a feature.

A guessed correction is precisely the claim the band exists to avoid. It would turn an honest "somewhere in this range" into a confident wrong number, and it would break the one thing that does work — a corrected reading is no longer comparable with the same device's reading from last week, because the guess can change underneath it.

There is a legitimate version of this, which is correcting against something you have actually checked the device against. That is what a manual offset is for, and it belongs on a page of its own about how to calibrate a phone decibel meter rather than being applied silently here.

What ±10 decibels actually costs

The band sounds tolerable until you convert it into the thing people use a level for. Allowed exposure time halves for every 3 dB, so ten decibels is a factor of ten either wayNIOSH.

A reading of 94 dBA is, at this precision, consistent with a room you could stay in most of the working day and one you should leave within minutes. That is not a rounding problem. It is the reason this page refuses to put a safe time next to a browser reading, and the reason the meter draws its uncertainty as a band on the dial rather than a needle at a point.

What the reading is genuinely good for

Comparison on one device.

Most of the error is systematic rather than random. Your microphone is wrong in a consistent direction, so the error largely cancels when you carry the same phone from the living room to the kerb, check the same machine before and after a repair, or compare Tuesday night with Wednesday night. The difference between two readings is worth far more than either reading on its own.

That is not a consolation prize. Most real questions are comparative: is this room louder than that one, is it worse than it was last month, did the new extractor fan help. A ±10 dB instrument answers all three, and a calibrated one answers them no better if you only ever take one reading.

It is also enough to tell an 85 dBA room — the level NIOSH recommends protection aboveNIOSH — from a 60 dBA one, which is the question most people arrive with. Twenty-five decibels is far outside any plausible error.

What NIOSH says apps are for

CDC puts it about as plainly as an agency can. Where a sound level meter is not available, sound measurement apps can provide a measure of area noise but may not comply with regulatory requirementsNIOSH.

The same page names the settings NIOSH recommends for a sound level meter: a 3 dB exchange rate, A frequency weighting, and Slow responseNIOSH. Those are the settings this meter runs — the dBA weighting it defaults to, the slow time response it defaults to, and the 3 dB rate behind its dose figures. The other curves are offered beside it, and selecting one withdraws the dose rather than restating it. Getting the settings right is not the same as being an instrument, but it does mean the number is the same kind of number the guidance is written about.

Why two phones disagree, and what to do about it

The spread in the study was between apps on one platform. Across handsets it is wider still, because the microphone, its housing, and whatever the operating system does before the audio reaches a browser all differ. Two people standing in the same room with different phones can reasonably see numbers several decibels apart, and neither of them is doing anything wrong.

The practical rule that follows is narrow and worth keeping. Compare readings from the same device, taken the same way, from the same position. Do not compare your reading with someone else's, and do not compare either with a figure from a table of typical values — published figures are ranges for a category of thing, recorded at a distance nobody statesNIOSH.

If you genuinely need the absolute number rather than the difference, there are three routes and only one of them is free. You can check the device against a source you already trust and carry a fixed offset from then on, which is the subject of how to calibrate a phone decibel meter. You can put a proper microphone in front of it, which fixes the input and none of the rest. Or you can use hardware built for the job, which is the only one of the three that produces a figure anybody else has to accept.

What it cannot be used for

An occupational compliance claim, a legal filing, an insurance report, or a medical decision. This page does not conform to IEC 61672, and no browser page does. If you need a figure that survives being challenged by someone with an interest in challenging it, you need Type 1 or Type 2 hardware and the tradeoffs belong under online vs hardware decibel meter.

If the argument you are building is with a neighbour or a landlord, the useful thing is not one impressive number. It is a neighbor noise log: dated readings taken the same way, from the same spot, with the same device, over enough nights that the pattern is the evidence rather than the level.

Checking whether your own reading is off

Three things are worth watching, and all three are visible without any equipment.

If the meter warns that automatic gain control could not be switched off, the number is being flattened — loud passages will read lower than they are and the session's average will be too kind.

If the reading flattens out at a loud event and stops responding to things getting louder, the input has run out of headroom. Treat the figure as a floor.

And if the number sits implausibly low in a room you have to raise your voice in, trust your voice. At 85 dBA you generally have to lift it to be heard three feet awayNIOSH; that check needs no hardware and does not drift.

The online decibel meter on this site shows the first two of those as warnings on screen rather than leaving you to infer them. Reading it well is mostly a matter of knowing what it cannot see.