[Advanced] — using the FFT on real, finite, noisy data, and the two problems that separate a clean spectrum from a misleading one.

The kid version first

In theory the FFT hands you a signal’s exact frequency content. In practice you only ever have a finite chunk of a signal, and that chunk lies to you in two specific ways: it smears sharp frequencies (leakage), and it forces a trade between knowing what frequencies are present and when (the resolution trade).

Spectral analysis is the craft of getting a trustworthy picture of a real signal’s frequencies despite both.

Leakage — the first problem

The FFT assumes your N samples are exactly one period of a signal that repeats forever. They almost never are — you grabbed an arbitrary window, so its start and end don’t line up. That discontinuity, when the FFT imagines the chunk repeating, injects fake frequencies:

   a clean tone, windowed at a non-integer number of cycles:
   the ends don't match → the FFT sees a "jump" → energy SMEARS
   across neighbouring bins instead of one clean spike

Spectral leakage is that smearing: a single true frequency spreads into a hump across many bins, burying weak nearby frequencies and making the spectrum look noisier than it is. It’s an artefact of the finite window, not the signal.

Windowing — the fix

A window function tapers the chunk smoothly to zero at both ends before the FFT, so there’s no discontinuity when it “repeats.” Multiply your samples by a window, then transform:

   samples × window  →  FFT
   (rectangular = no window = worst leakage)
WindowCharacter
Rectangular (none)Best frequency resolution, worst leakage
Hann / HammingThe everyday default — good leakage suppression, mild resolution cost
Blackman / Blackman–HarrisVery low leakage, wider main lobe (worse resolution)
Flat-topPoor resolution, but accurate amplitude — used for calibration

The universal trade: every window suppresses leakage (lower side-lobes) at the cost of resolution (wider main lobe). A better main-lobe/side-lobe trade is impossible to escape — you’re choosing how to spend it. Use Hann by default; reach for others only with a specific reason (Blackman for weak signals near strong ones, flat-top when you need the amplitude to be exact).

The resolution/latency trade — the second problem

Frequency resolution is fs / N: to distinguish two close frequencies you need a large N, which means a long time window. But a long window blurs when things happened. You cannot have fine frequency resolution and fine time resolution at once.

   long window   → sharp frequencies, but "somewhere in this long span"
   short window  → precise timing, but blurry frequencies

This is the time-frequency uncertainty principle — the DSP sibling of Heisenberg’s, and the same mathematics. It’s not an engineering limitation you can design around; it’s fundamental. Every spectral tool is a choice of where to sit on this trade.

The spectrogram — seeing frequency change over time

A single FFT gives one spectrum for the whole chunk — useless for a signal whose content changes (speech, music, a bird call). The Short-Time Fourier Transform (STFT) fixes this: slide a window along the signal, FFT each position, and stack the results into a spectrogram — time on one axis, frequency on the other, colour for magnitude:

   freq ▲  ████░░  ← a rising tone sweeps up over time
        │  ░░████
        └──────────► time

The spectrogram is the workhorse visualisation of DSP — it’s how you see speech formants, music notes, a modem’s tones, a machine’s changing vibration. And the window length is exactly the resolution/latency trade made visible: a short window gives crisp timing and blurry pitch, a long window the reverse. Speech analysis uses short windows (timing matters); tuning a note uses long ones (pitch matters).

Mel-spectrograms — a spectrogram warped to the perceptual mel frequency scale — are the standard input to speech and audio ML models, which is a direct bridge from DSP to audio ML.

Practical spectral analysis

Turning an FFT into an honest spectrum:

  1. Window (Hann, usually) before the FFT
  2. Zero-pad if you want a smoother-looking spectrum — but understand it interpolates, it does not add real resolution (that needs more actual samples). A common misconception
  3. Average multiple windows (Welch’s method) to reduce noise variance in the estimate — trading some resolution for a cleaner, more reliable spectrum. The standard for estimating a noisy signal’s power spectrum
  4. Convert to dB (log scale) to see weak and strong components together — the ear and eye are logarithmic
  5. Label the axis correctly — bin k = k·fs/N Hz
from scipy import signal
f, t, Sxx = signal.spectrogram(x, fs, window='hann', nperseg=1024)
# f: frequencies, t: times, Sxx: power at each — plot as an image

Key insight

A finite chunk of signal lies about its spectrum in two fixed ways — leakage (smearing from the window’s edges) and the resolution/latency trade (you can’t pin down frequency and time at once) — and spectral analysis is the craft of managing both. Windowing (Hann by default) buys leakage suppression at a resolution cost you can’t escape, and the spectrogram makes the time-frequency trade visible by sliding short FFTs along the signal. The uncertainty principle here isn’t a limitation to engineer around; it’s the terrain.

Source: [reference] — Aug 2026.

↑ Back to Digital Signal Processing