Sound Spectrum Analyzer

A sound level meter answers one question: how loud. It is a good question and it is not always the one you have. A fridge compressor, a neighbour's subwoofer and a laptop fan can all read the same on a single number and none of them is the same problem.

This page answers the other question. What is actually in the sound, where in the frequency range it sits, and whether it is steady, intermittent, or one thing that happened once.

Three views, three questions

Spectrum is what is there right now — level against frequency, across the range human hearing covers. It is the view that finds a tone: a hum, a whine, a ring, anything with a definite pitch shows as a spike where broadband noise shows as a hill.

Spectrogram is the same information over time, with frequency up the side and level as brightness. It is the only one of the three that can tell a steady noise from an intermittent one, because a spectrum has already thrown the time axis away. A fan is a horizontal stripe. A door slam is a vertical one. A neighbour's music is a horizontal stripe that stops.

Octave splits the range into the bands acoustics work is actually reported in, and is the view to reach for when you want to write something down rather than look at it.

Where the sound comes from

Three inputs, and the choice changes what the numbers are about.

The microphone is the room you are standing in, heard through a device whose frequency response nobody published. Everything the rest of this site says about uncertainty applies to it.

System audio shares a browser tab and reads what it is playing. Pick a tab and tick Share tab audio: without that box you get a picture and silence, and a whole-screen share usually carries no audio track at all. The picture is dropped the moment the share starts, because this page has no use for it and no reason to leave a capture indicator lit.

An audio file is decoded in your browser and played through the same analyser, with play, pause, stop and a mute for the monitor. Muting silences the speakers and not the analysis: the plot and the figures go on reading the file. Nothing is uploaded. The decoding runs on your own machine and the file never leaves it.

The last two are not the room, and that is worth saying out loud rather than burying. A file is a recording made by equipment you may know nothing about. A tab is whatever a website decided to play, at whatever gain it chose. Neither has a sound level in the sense the meter means. What the analyser can tell you about either is shape, which is what it was for anyway.

Finding the hum

The single most useful thing this page does is name a hum.

Electrical hum sits at the mains frequency — 50 Hz in most of the world, 60 Hz in North America — and at its multiples. Point the analyser at a buzzing transformer, a dimmer, a fridge or an amplifier and you will see a spike at the mains frequency with smaller ones stacked above it at two, three and four times that. The dominant frequency readout names it outright, with the nearest musical note beside it.

That matters because the frequency tells you what the thing is in a way the level never does. A 50 Hz spike with harmonics is electrical. A broad hill somewhere between 100 and 300 Hz is usually mechanical — a motor, a pump, a compressor. A hiss with no structure at all is air, either moving through a vent or being made by the microphone itself.

The note readout is a tuner good enough to name a string or separate a 50 Hz hum from a 60 Hz one. It is not a calibrated frequency counter, and it will report nonsense when there is no tone present to report — which is most rooms.

Why the levels here are relative

The meter on this site publishes an absolute figure with a band around it: an uncalibrated estimate, typically within about ±10 dB of a reference instrumentKardous & Shaw. This page does not repeat that claim per frequency, and its axis is deliberately not labelled in absolute terms.

The reason is that a device microphone's error is not one number, it is a curve. Its response is uneven across the range, and it is unevenly uneven from handset to handset — which is precisely the thing a spectrum is drawing. A broadband level can be corrected with a single offset; a spectrum cannot.

So read the plot for shape, and read the online decibel meter for level. Where the sound sits, which peak is the loud one, whether the hum got better after you moved the fridge — all of that survives an unknown microphone response, because the same response is under every point on the curve.

Two things on this page are not relative, and neither of them is the plot. The measurement session below it reports broadband levels, worked out the same way the meter works out its number and carrying the same band around them — they are here because a reader who came to look at a hum should leave with figures they can quote, not because a spectrum has earned an absolute axis. The signal panel is absolute about something else entirely, and has its own section next.

The signal figures

Four figures under the plot describe the signal rather than the room: peak, RMS, crest, and a clip lamp.

Peak is the loudest single sample in the quarter second just gone, and RMS is the energy average over the same window. Crest is the gap between them, and it is the one that says what kind of sound this is. A steady tone crests about three decibels over its own average. Speech and music run ten to twenty. A signal that crests at almost nothing is either a square wave or something already squashed flat by a limiter. The lamp lights if any sample in the window reached the top of the range, which is the one condition that makes everything else on the page unreliable at once.

Peak and RMS are in dBFS: decibels relative to full scale. Full scale is the largest number the input can represent, so zero dBFS is the top of the file or the top of the converter, and every level below it is a ratio to that. It is the one absolute figure this page can carry honestly, because it is a fact about the numbers rather than a claim about the air. No microphone sensitivity goes into it, no distance, no reference pressure.

Which is exactly why dBFS is not dB SPL. A signal six decibels under full scale is using half the input's range, and that tells you nothing about how loud anything was: record the same room twice with the gain moved and you get two dBFS readings and one sound level. Read dBFS to find out whether the input is set sensibly — near the top it will clip, far below it you are looking at the noise floor. Read the online decibel meter for how loud.

Two channels, when there are two

A stereo file or a shared tab has two channels, and the analyser will show either of them, the two summed, or the difference between them. Left and right are the channels as they arrived. Mid is the two added and halved, which is the mono fold-down and what a single speaker will play. Side is the difference: everything the two channels disagree about, silent on a mono recording and loud on a wide one.

Beside them is a correlation figure between minus one and plus one. It reads plus one when the channels agree, near zero when they carry unrelated material, and minus one when one channel is the other inverted. That last case cancels to silence the instant anything sums them, so a recording sitting near minus one is one that will vanish on a phone speaker. You can watch it happen: switch to Mid and the plot empties.

The controls appear only when the source really has two channels, which the microphone does not on almost any device. Splitting a mono signal into two identical halves would read plus one for ever and describe the wiring rather than the sound, so on a mono source the page says so and leaves the buttons off.

Mid and side are a view of a signal, not of a room. Two microphones a metre apart would give a mid/side pair that said something about a space; a file and a browser tab give one that says something about a recording.

Octave and third-octave bands

An octave is a doubling of frequency. Splitting the range at octaves gives ten bands from 31.5 Hz to 16 kHz, which is few enough to hold in your head and is what the meter's own spectrum shows.

Third-octaves split each of those into three, giving thirty-one bands from 20 Hz to 20 kHz. That is the resolution building-noise assessments, room acoustics and equipment specifications are written in, and it is fine enough to separate a hum from the noise either side of it while still being a table rather than a picture.

Use octaves to orient yourself and thirds to write something down. Both are exported as CSV, with the sample rate, the window and the weighting in the header, because a column of levels is unreadable six months later without them.

The measurement session

Underneath whichever view is on screen, the analyser can run a timed session and report six figures.

Leq is the energy average over the session — the single number that represents the whole span. Lmax and Lmin bound it. L10, L50 and L90 are the levels exceeded a tenth, half and nine tenths of the time, and the gap between L10 and L90 is how variable the noise was: a wide gap means events, a narrow one means a steady source.

Those last three are the figures noise assessments quote, which is why they are here and not held back for a paid tier. Thirty seconds is enough for a steady source; five minutes suits an office or a plant room; fifteen is the length environmental surveys are usually specified at.

FFT size and smoothing

The window length is a trade and there is no setting that wins both sides of it. A long window resolves frequency finely and blurs time, so a 50 Hz hum separates cleanly from 60 Hz but a door slam smears. A short window does the reverse.

The default sits in the middle. Move it up when you are chasing a tone and down when you are watching something change. Smoothing is a display control only — it steadies the picture between frames and does not touch the measurement session's figures.

Telling the microphone from the room

Some of what you see is not in the room. A device microphone makes its own noise, and on a spectrum it has a recognisable shape: a broad hiss rising towards the top of the range with no structure in it, sitting at whatever the input's floor happens to be.

The way to separate the two is to look at what changes. Cover the microphone with a hand, or take the device somewhere quieter, and watch which parts of the curve move. Anything that drops was the room. Anything that stays put is the device, and no amount of staring at it will make it a measurement of anything.

This is also why a very quiet room is the hardest thing on this page to read: below the input's own floor there is nothing left to draw, and the curve flattens into the hardware rather than into silence.

Which curve to analyse on

The A, C and Z buttons do here what they do on the meter, applied to every frequency rather than collapsed into one number, which is the clearest place on the site to see what a weighting curve actually is: switch from A to C and watch the bass return.

Use Z when you want the signal as it arrived, which is usually the right choice for hunting a hum — A-weighting is designed to discount exactly the low frequencies a hum lives in. Use A when you want the shape that corresponds to the figures health agencies publish. The difference between them is set out under dBA vs dBC.

What this page deliberately does not do

It still puts no absolute figure on the level axis of the plot, for the reason set out above. The frequency axis is labelled, because a frequency is a number the FFT actually knows. The signal panel's dBFS is the one absolute reading here, and it is absolute about the input rather than about the room.

It does not treat mid and side as a picture of a room, and it will not draw a stereo view of a mono source. It does not record what it hears, upload the file you open, or keep anything once the tab closes.

And it calibrates nothing. A file has a level relative to full scale and no level in decibels of sound; a tab has whatever gain the site chose. The decibel chart is where published sound levels live, and the meter is where a reading with a stated band around it comes from.