How to Read a Spectrogram

A spectrogram looks like a weather map, and most people stare at one for a while without knowing what they are supposed to notice. There is less to learn than it appears: three axes, half a dozen shapes, and one common misunderstanding. After that you can read most of what you will run into.

Every picture on this page is a real spectrogram of a clip you can open in the analyzer yourself. Nothing here is an illustration.

The three axes

A spectrogram of a sung note, three seconds long, showing stacked horizontal lines — Three seconds of a sung note. The same picture the analyzer draws when you click the Sung note example.

Left to right is time. The same as a video timeline. The clip above is three seconds long.
Bottom to top is pitch. Low sounds sit near the bottom, high sounds near the top.
Brightness is loudness. Black is silence; yellow and white are loud.

That is the entire language. Everything after this is learning what ordinary sounds look like once they are drawn this way.

One thing worth knowing early: the vertical axis in these pictures is logarithmic, meaning each equal step upward doubles the frequency. Musical hearing works the same way, so a log axis spaces notes evenly. On a linear axis, everything musical is crushed into the bottom few percent of the image.

Six shapes worth recognising

1. A held note: one line, with more stacked above it

Evenly stacked horizontal lines with a gentle wobble, from a sung note — A single sung note. The lowest line is the note; the lines above are its overtones.

A voice or an instrument holding a pitch draws a stack of horizontal lines. The lowest one is the note itself. The ones above belong to the same note — see the next section, because this is where most people go wrong.

The gentle waviness is vibrato: the singer's pitch drifting slightly up and down, roughly five times a second. The higher lines wobble more than the low one, which is normal — they move by the same percentage, and a percentage of a big number is a bigger distance.

2. Mains hum: a hard, straight line near the bottom

A straight bright line low in the image with further lines evenly above it — Hum from mains power: a rigid line at 50 Hz with harmonics above it, running the whole width.

Electrical hum draws a very straight line at 50 Hz, or 60 Hz in North America and parts of Asia, with more lines at exact multiples above it. It runs the full width of the picture because it never stops and never varies.

If you see this, it is almost never the performance. It is the wiring, a ground loop, or a power supply sitting too close to a cable. Point at the lowest line in the analyzer: if it reads 50 or 60 Hz, that is your answer, and a narrow cut at that frequency usually fixes it.

3. A noise floor: haze with no lines in it

A soft textured haze filling the upper part of the image above the signal — Hiss and room tone: texture without structure, filling the space around the content.

Preamp hiss, air conditioning and room tone have no pitch, so they fill the picture with a soft texture instead of lines. It is usually strongest towards the top, which is why a noisy recording sounds fizzy rather than rumbly.

What matters is not how bright the haze is on its own, but how far it sits below the content. If the haze is nearly as bright as the voice, the recording is in trouble and no amount of processing will fully rescue it.

4. A click: a thin vertical stripe

Narrow vertical stripes running the full height of the image — Clicks. Something abrupt happens at every frequency at once, so it draws a full-height line.

Anything sudden — a mouth click, a chair, a dropout, a bad edit — contains every frequency at once, so it draws a vertical line through the entire height.

This shape is the best argument for opening a spectrogram at all. A click can be two milliseconds long. You can hear that something happened but you will never find it by scrubbing back and forth. Here you can see precisely where it is, and the time readout tells you the moment to cut.

5. Short high strokes: birdsong, keys, consonants

Short curved bright strokes high in the image — Birdsong. Brief sounds that also change pitch draw short curved strokes near the top.

Sounds that are short and change pitch draw little curved strokes. Birdsong is the clearest case, but the "s" and "t" in speech, keys jingling and a tambourine all look related: narrow, bright, and high in the picture.

If a voice recording looks harsh, this is the region to inspect. Sibilance shows up here as repeated bright flecks every time the speaker hits an "s".

6. A pitch that moves: a line that travels

A single bright line climbing from the bottom left, steeply at first then flattening — A tone sliding from 20 Hz to 20 kHz over ten seconds.

When a pitch slides, its line slides with it. A siren, a slide guitar, a synth sweep and tape wobble are all this shape at different speeds.

The mistake everyone makes

A stack of evenly spaced lines is usually one note, not several.

This is the single most common misreading, and it catches experienced people too. Almost every instrument and every voice produces a fundamental frequency plus overtones at two times, three times, four times that frequency and upward. They are all part of one sound. You hear one pitch; you see a dozen lines.

The way to check takes two seconds. Leave Mark lines of the same note ticked in the analyzer and point at the lowest bright line. Dashed markers appear at 2×, 3×, 4× and 5×. If they land on the lines above, those lines belong to the note you are pointing at — it is one sound, not five.

If a line does not line up with the guides, it belongs to something else: a second instrument, a hum, or noise. That is the moment the picture starts answering questions instead of posing them.

What the two settings actually change

Two controls change what you can see. Both are trade-offs, and neither setting is correct in general.

FFT size: sharp in time, or sharp in pitch

One wide smear where two clicks occurred — FFT 2048 — a 46 ms window, so two clicks 25 ms apart merge

Two distinct narrow vertical stripes — FFT 512 — an 11.6 ms window separates them

Both pictures are the same third of a second of the same file — two clicks, 25 milliseconds apart.

The analyser works by looking at a short window of audio at a time. A long window can tell pitches apart finely but cannot say exactly when anything happened. A short window pins down timing but smears pitch. That is a law, not a limitation of this tool: you cannot have both at once.

So: lower the FFT size when you are hunting for clicks and edits. Raise it when you are trying to separate two close notes. Zooming in does not help with timing, because the window length is what sets it.

Linear or logarithmic frequency

A sweep drawn on a logarithmic axis, curving — Logarithmic — musical spacing; the bass is readable

The same sweep drawn on a linear axis, a straight diagonal — Linear — equal hertz per pixel; everything musical is at the bottom

The same ten-second sweep, drawn both ways.

Use logarithmic for anything musical. Notes are spaced by ratios, so a log axis puts an octave the same distance apart wherever you are, and the bass — where most of the trouble lives — gets real estate instead of a sliver.

Use linear when you care about exact hertz in the top half: checking where a low-pass filter starts, or looking at harmonics of a fixed tone, where even spacing is easier to read.

Now open one yourself

Reading about shapes only goes so far. Open the analyzer and click the three examples in order — birdsong, a sung note, then the noisy recording — and find the shapes above in each one. It takes about two minutes and it is the part that makes the rest stick.

Two things to try while you are there:

Point at the lowest bright line in the sung note. The readout names the note and its octave, and the dashed same-note markers show you which of the lines above belong to it.
In the noisy recording, point at the hard line near the bottom. It reads around 50 Hz — and now you know what to cut.

Next: which sounds live in which frequency band, with a note-to-frequency table.