How to Read a Spectrogram

By VocalCut team · 2026-10-21

To learn how to read a spectrogram, open it in the free Spectrogram tool and start with the axes: time runs left to right, frequency runs bottom to top, and brightness shows how loud each frequency is at each moment. After that, reading one is pattern recognition, because a handful of shapes (lines, stacks, stripes, haze) account for almost everything you'll see. The theory is covered in what a spectrogram is; this guide sticks to the practical side.

Set up the tool in under a minute

Open the Spectrogram and drop in a file. MP3, WAV, FLAC and M4A work in most browsers (anything your browser can decode). The audio is decoded in your browser, mixed down to mono, and analysed in a background worker. You'll see "Analysing…" with a percentage while it runs. It's free, there's no account, and your audio never leaves your device.
Hover over the picture. A readout under the image shows the time, the frequency, and the level in dB relative to the loudest point in the file (so 0 dB is the peak and −40 dB is 40 dB quieter than it). Points too quiet to measure say "silent". This readout is your measuring tape for everything below.
Click anywhere to listen. Playback starts from the point you clicked, and a thin white playhead moves across the image so you can match what you hear to what you see. The Play button starts from the beginning.
Adjust the four controls:
- Frequency scale: Logarithmic (the default, starting at 20 Hz) or Linear.
- Window size: 1024 · sharper timing, 2048 · balanced (default), 4096 · finer pitch, 8192 · finest pitch.
- Dynamic range: a slider from 40 to 120 dB (default 90) that sets how far below the loudest point is still drawn.
- Colours: Ice, Amber or Greyscale.
Save it with Download PNG if you want to keep or share the image. The axes are drawn in, so the PNG makes sense on its own.

Only a window-size change re-runs the analysis. Scale, range and colour redraw instantly, so flip between them freely.

Reading a spectrogram: the shapes that matter

Nearly every recording is built from the same few shapes. Once you can name them, a dense picture becomes a list of things you can check by ear.

Horizontal lines: sustained tones

A held note, a whistle, a sine tone or a feedback squeal draws a horizontal line at its pitch: length is duration, height is frequency, brightness is level. Hover over it to read the exact pitch, handy for finding a ringing resonance.

Harmonic stacks: voices and instruments

Most real sounds aren't one line but a stack of them. A sung or played note has a fundamental (the lowest line, the pitch you hear) plus harmonics at whole-number multiples above it. A 200 Hz note has lines at 400, 600, 800 Hz and so on, evenly spaced in hertz. When the pitch moves, the whole stack bends together, which is how you pick out a melody or a speaking voice.

On the default logarithmic scale, those lines look like they bunch closer together as they rise. They haven't moved. A log axis gives every octave the same height, and the gap from 200 to 400 Hz is a full octave while the gap from 1,000 to 1,200 Hz is much less than one. Switch to Linear and the spacing evens out. Audio frequencies explains why pitch works in ratios like this.

Vertical stripes: transients

Anything sudden (a drum hit, a hand clap, a plosive "p", a click) spreads energy across many frequencies at once, so it draws a vertical stripe. The shape tells you what it is:

A click or digital glitch is a razor-thin line that runs almost the full height of the plot.
A kick drum is a short, wide, bright blob at the bottom, often with a fainter tail above it.
A snare or clap is a stripe concentrated in the mids and highs, with a noisy smear rather than clean lines.

Hum: a thin line that never stops

Electrical hum is a thin, steady horizontal line at 50 Hz or 60 Hz (depending on the mains frequency where it was recorded), usually with fainter lines at multiples like 100/150 Hz or 120/180 Hz. The giveaway is that it runs the entire length of the file, straight through pauses and silences, and never changes pitch. Hover over it: if the readout sits right at one of those values, you've found it.

Sibilance: bright bursts up high

Sharp "s", "sh" and "ch" sounds show up as bursts of bright energy roughly in the 4–10 kHz region, lining up with the consonants in the words. Bursts that are much brighter than the voice around them are a sign of harsh sibilance.

Hiss: a dim haze across the top

Broadband hiss from a noisy preamp or a cheap mic looks like a faint, even haze across the upper part of the plot, present everywhere, including between words. Raising the Dynamic range makes it easier to see.

The lossy ceiling: a flat edge near the top

MP3 and AAC files often cut off sharply above a certain frequency, leaving a dark band across the top with a perfectly straight edge. Where that edge sits depends on the encoder and bitrate. Bitrate vs sample rate explains the trade-offs behind it.

Log or linear: which scale to use

Use Logarithmic most of the time. It matches how we hear pitch, where every octave feels like the same size step, and it gives the bass room to breathe. On a linear scale, everything from 20 Hz to 1 kHz squeezes into a sliver at the bottom, even though that's where kick drums, bass and the fundamentals of voices live. Hum at 50 or 60 Hz is easy to read on log and nearly invisible on linear.

Switch to Linear when you care about the top half: checking for a codec ceiling, inspecting high-frequency detail and hiss, or confirming that a stack is evenly spaced harmonics (it will look like a ladder with equal rungs). If you need a refresher on which frequencies are even audible, see the audible frequency range.

Window size: timing versus pitch

The window size is how many samples go into each FFT (Fast Fourier Transform) slice, and it works as a trade-off rather than a quality setting. A bigger window measures frequency more finely but averages over more time, so sharp events smear sideways. A smaller window catches transients crisply but blurs pitch vertically. You can't have both at once; every FFT-based spectrogram works this way.

The frequency resolution is simple to work out: the sample rate divided by the window size. At 44.1 kHz:

Window	Frequency step	Time per window	Good for
1024	≈ 43 Hz	≈ 23 ms	Drums, clicks, plosives, edit points
2048	≈ 21.5 Hz	≈ 46 ms	General listening, speech, mixed music
4096	≈ 10.8 Hz	≈ 93 ms	Bass notes, hum, tuning checks
8192	≈ 5.4 Hz	≈ 186 ms	Separating very close low pitches

If drums look like blurry columns, go smaller. If the bass is one fat smudge instead of distinct lines, go bigger.

Dynamic range and colours

Dynamic range sets the floor: anything quieter than that many dB below the loudest point is drawn as black. Lower it (say, 50–60 dB) to hide the noise floor and see just the main content clearly. Raise it toward 120 dB to reveal quiet things like reverb tails, fade-outs, faint hiss or a buried hum. If the picture looks like a washed-out fog, lower it. If it looks mostly black, raise it.

Colours don't change the data. Ice is the default. Amber gives a warmer contrast that some people find easier on dim detail, and Greyscale is best for printing or for anyone who finds colour gradients hard to tell apart. Brighter always means louder.

Worked example: diagnosing a voice memo

Say you load a two-minute voice memo recorded at a desk. Here's what reading it might turn up:

Speech harmonics. Wavy stacks of lines in the lower mids, rising and falling with the voice, separated by short dark gaps between phrases. That's the content you want to keep.
A hum line. A thin line near the bottom that keeps going through every pause. The hover readout says 60 Hz, with a fainter line at 120 Hz. A charger or other mains-powered gear nearby is a common source.
An "s" burst. Bright flares up high on the "s" sounds, noticeably hotter than the rest of the voice.
A mouth click. A hairline vertical stripe from bottom to top, between two words. Click just before it to hear it.

A spectrogram diagnoses; it doesn't fix anything. For steady hum and hiss, try the noise reducer; its Music mode has an optional 50/60 Hz hum notch, and how AI noise reduction works explains what it can and can't remove. For harsh sibilance or a ringing resonance, a narrow cut with the equalizer at the frequency you measured is the usual move. How to EQ a vocal walks through that. If the waveform was flattened at the top, that's a different problem: see what audio clipping is.

Troubleshooting

The picture is all dark, or a washed-out fog. Adjust Dynamic range. Mostly black means the range is too small for a quiet recording, so raise it. Uniform haze means the noise floor is being drawn, so lower it.

Everything is crammed at the bottom. You're probably on Linear. Switch back to Logarithmic so the bass and mids get room.

Drum hits look smeared sideways. The window is too large for transients. Drop to 1024.

A long file looks squashed and detail is lost. The whole file fits the width of the image, so an hour-long recording gives each pixel column several seconds. Trim the section you care about with the audio cutter and load that instead.

A stereo problem doesn't show up. The tool analyses a mono mixdown, so a fault on one side gets averaged with the other, and anything out of phase between the channels can partly cancel. Split the file into left and right with the channel splitter, then load each channel on its own.

Frequently asked questions

What do the colours on a spectrogram mean? The colour or brightness of each point shows how loud that frequency is at that moment. Dark means quiet or silent, and brighter means louder, with the brightest areas closest to the loudest point in the file. In Vocal Cut's Spectrogram, the Ice, Amber and Greyscale options only change the palette, not the data, and the Dynamic range slider sets how far below the peak is still drawn before it turns black.

Which window size should I use for a spectrogram? Start with 2048, which balances timing and pitch detail for most music and speech. Drop to 1024 when you're looking at drums, clicks or edit points, since smaller windows show sharp events more precisely. Go up to 4096 or 8192 for bass notes, hum or tuning checks, where you need fine frequency detail and don't mind timing getting blurrier. The trade-off is unavoidable: more pitch detail always means less timing detail.

Why do harmonics bunch together near the top of a spectrogram? Harmonics sit at whole-number multiples of the fundamental, so they're evenly spaced in hertz. A logarithmic frequency scale gives each octave the same height, though, and the higher harmonics are less than an octave apart, so they get drawn closer and closer together as they rise. Switch the view to a linear scale and the same harmonics appear as an evenly spaced ladder.

Can a spectrogram show if an MP3 was converted to WAV or FLAC? Often, yes. Lossy encoders like MP3 usually discard content above a cutoff frequency, leaving a dark band with a sharp, straight edge near the top of the spectrogram. Converting that file to WAV or FLAC doesn't bring the missing frequencies back, so the edge stays. A genuinely lossless recording usually fades out gradually instead. It's strong evidence rather than proof, since some recordings are naturally dull at the top.