How to Read an Audio Spectrogram (And What It Hides) | Rohan
Rohan Vijayakumar explains how a spectrogram maps audio energy across time and frequency, how to read harmonics and window-size tradeoffs, and the three things a spectrogram cannot tell you.
A spectrogram is the picture audio people stare at when a waveform isn't enough. Rohan Vijayakumar teaches what that picture actually measures, when to use it, and the three things it will not tell you, even when it looks decisive.
Waveform vs. spectrogram
A waveform plots air pressure against time, so you can see loud and quiet, but you can't see which pitches are present. Two completely different sounds can share a similar-looking wiggle. A spectrogram exists to unpack that wiggle into pitches, turning a single line into a map of energy over time and frequency.
Reading the axes
Time runs left to right, frequency runs up, and each cell's brightness shows how much energy sat in that pitch at that moment. A sung note lights up a band. Noise looks like speckle with no pitch to hold onto. Read the picture as a histogram of loudness sliced by time and by frequency, not as a single measurement.
Harmonics and spectral brightness
A sung note is not a single line on the spectrogram, it's a stack, a lowest band with fainter copies above it. Those copies are harmonics. Brightness is energy, not correctness, so a bright smear can mean a distortion artifact or two notes fighting each other. The point is to read the shape of what's there, not to assume brightness alone tells you what it means.
The window size tradeoff
A spectrogram is built from a sliding window, and the window's length trades one kind of clarity for another. A short window is sharp in time, showing you the exact attack of a drum hit. A long window is sharp in pitch, showing you a sung note sitting still. You cannot have both at once. If a section of the picture looks blurry, check the window size before assuming the problem is in the recording itself.
What a spectrogram hides
Three things don't show up no matter how good the picture looks. Phase: two sounds can produce the same spectrogram and still cancel each other out in the air. Overlapping voices: multiple sources smear into a single stain instead of separating cleanly. Quiet sounds: consonants and other low-energy detail can fall below the visible floor entirely. Use a spectrogram to find a note, a breath, a cut, or a dropout, not to declare a mix clean or to claim a model understood a lyric.
Key takeaways
- A spectrogram maps energy across time and frequency; a waveform only shows loudness over time.
- Harmonics appear as a stack of bands above a fundamental pitch, not as a single line.
- Window length trades time precision against frequency precision, you can't maximize both at once.
- A spectrogram cannot show phase, cannot cleanly separate overlapping voices, and cannot reveal sound below its energy floor.
- Use a spectrogram to locate structure (a note, a breath, a cut), not as proof that a mix is clean or a lyric was understood.
Who this is for
Anyone working with songs, speech, or lyrics who relies on spectrograms, and wants a clear sense of what the picture is actually showing versus what it only looks like it's showing.
Chapters
Full transcript(auto-generated, with timestamps)
Waveforms vs Spectrograms: Unpacking audio
[0:00]Hi, this is Rohan Vijay Kumar from Humanitarians.ai. A spectrogram is the picture audio people stare at when a waveform is not enough. Today, I will teach you what that picture is measuring, when to use it, and the three things it will not tell you even when it looks decisive. If you work with songs, speech, or lyrics
Reading time, frequency, and energy axes
[0:22]Tools, you will meet this picture. It is a map of energy over time and pitch. Use it to see structure. Do not use it as proof of what the sound is or that a model understood the lyric. Start with the picture most people already know, a waveform. It plots air pressure against time. You can see loud and quiet. You
Understanding harmonics and spectral brightness
[0:42]Cannot see which pitches are present. Two totally different sounds can share a similar wiggle. That is why audio people switch to a spectrogram. It unpacks the wiggle into pitches. I am going to ask Claude to draw the map the honest way. No rainbow heat palette. Just the two axes and energy as brightness. Watch the
Sliding windows: Time vs pitch sharpness
[1:03]Ask, then the result. Time runs left to right. Frequency runs up. Each cell is how much energy sat in that pitch at that moment. A sung ah lights a band. Noise is speckle with no pitch to hold. You are looking at a histogram of loudness sliced by time and by
What it hides: Phase, overlapping voices, and quiet sounds
[1:20]Frequency. A sung ah is not one line. It is a stack, a lowest band, then fainter copies above it. Those copies are harmonics. Brightness is energy, not correctness. A bright smear can be a acquired distortion or two notes fighting. Read the shape, do not baptize it. The picture is made with a sliding
Librosa Python exercise in Claude
[1:40]Window. A short window is sharp in time. You see the attack of a drum. A long window is sharp in pitch. You see a sung note sit still. You cannot have both at once. If a cut looks blurry, check the window before you blame the singer. What it hides? Phase, two sounds can share this picture and still cancel in the air. Overlapping voices smear into one stain. Quiet consonants fall under the floor. Use a spectrogram to find a note, a breath, a cut. Do not use it to decide a mix is clean or that a model understood the lyric. The split is practical. Use a spectrogram to find a note, a breath, a cut, a buzz, a dropout. Do not hand it the last word
On a clean mix, isolated stems, or understood lyrics. The picture is energy. Meaning is still on you. Remember the split. A spectrogram shows energy by time and frequency. Use it when you need to see structure. Do not hand it the last word on phase, on overlapping sources, or on anything quieter than the floor. Your turn. Paste this into Claude. Using Librosa, plot a spectrogram of a 440 hertz tone next to a two-note chord. In two sentences, what does the chord show that the tone does not? And what do both plots still hide? Run it on a clip of your own. Look for second band on the chord and for everything the magnitude plot refuses to say. What a spectrogram shows. Rohan Vijayakumar, lyrical literacy. Humanitarians AI.
More from Humanitarians AI Fellows
3:05Nyquist Limit, Sample Rate & Aliasing Explained | Rohan
4:29Simpson's Paradox: Why Aggregate Metrics Can Mislead You | Ushasvi
2:343 Ways Language Models Fail (And Why Context Pasting Fails) | Vedanshu
2:03Why Contributor Profiles Beat Brand Posts on LinkedIn | Ankita
2:12The 3-Act Volunteer Story System: Build a Repeatable Series | Ankita
4:04