How to read a spectrum analyzer, pre vs post

An audio spectrum analyzer shows frequency on the horizontal axis and level on the vertical axis. Peaks show where the signal has more energy; they do not automatically show a problem. Read the scale and analyzer settings first, then compare what the display suggests with what you hear.

Before changing the sound, check the FFT size, display tilt and smoothing. These settings can make the same audio look different. A smooth trace is useful for broad balance; a detailed trace is better for locating a narrow ringing frequency. The examples here use Vitric's Bevel EQ, whose analyzer can show the signal before the EQ, after it, or both together.

This article walks the mechanism: what one frame of the display actually measures, what each setting trades away, why pre and post taps answer different questions, and how to run a resonance hunt where the analyzer helps instead of decorates.

What one frame of the display actually is

An FFT analyzer grabs a block of samples, N of them, and estimates the signal's content at regularly spaced frequencies. For a real-valued signal, the non-redundant bins run from zero to half the sample rate. A 4096-point FFT at 48 kHz gives you bins spaced about 11.7 Hz apart. The analysis window and calibration also affect the displayed levels.

The catch is that the bins are spaced linearly and the display axis, like your hearing, is logarithmic. At 11.7 Hz spacing, the octave from 10 kHz to 20 kHz contains roughly 850 bins. The octave from 40 to 80 Hz contains three or four. So the top of the display is squeezing hundreds of bins into a few pixels, while the bottom is stretching a handful of bins across the widest, most contested region of the mix. Whatever precision an analyzer has, it has the least of it exactly where bass decisions get made.

FFT size: resolution against speed

A bigger FFT narrows the bins. It also lengthens the block: 4096 samples at 48 kHz is about 85 milliseconds of audio, so each frame of the display describes almost a tenth of a second, and the trace blends events across that interval, even if overlapping frames let the display refresh more often. A small FFT is the opposite. A 1024-point analysis updates quickly and feels snappy, but its bins are about 47 Hz wide, which means everything below 100 Hz lands in two or three bins. Separating nearby components at 55 and 80 Hz is difficult at that resolution. Interpolation may estimate an isolated peak between bins, but it does not remove the underlying tradeoff.

The practical rule: use a large FFT when you are inspecting low end, and treat the bottom octave of a fast, small-FFT display cautiously. Apparent low-end movement can reflect the display settings as well as changes in the bass itself.

Tilt: why flat does not look flat

Pink noise carries equal power per octave, rather than per hertz. Since each octave upward spans twice as many hertz, its power per fixed-width bin falls by about 3 dB per octave. Music can have a broadly falling spectrum too, but its arrangement and sound choices determine the shape. A good mix does not have to match pink noise.

To compensate, analyzers apply a tilt, commonly 3 to 4.5 dB per octave, so balanced program draws approximately level. This is a display convention, not a measurement, and different tools pick different slopes. If you do not know your analyzer's tilt, you do not know what "flat" on it means, and traces from two plugins with different slopes are not comparable. White noise, which has equal energy per hertz and sounds piercingly bright, is what actually draws flat on an untilted display. That alone should end the habit of mixing toward a level line.

Averaging and smoothing hide different things

These two settings get conflated because both make the trace calmer. They operate on different axes and they hide different information.

Averaging works across time: each drawn frame blends the last several FFT frames. Long averaging gives you a steady trace that reads like the tonal fingerprint of the material, which is exactly what you want when judging overall balance, and exactly what erases transients. A snare's crack does not survive the blend; only its sustain does.

Smoothing works across frequency: each drawn point blends neighboring bins, usually specified in fractions of an octave like 1/3 or 1/6. It produces the smooth, publishable kind of curve. It also blends narrow features with their neighbors, making a small resonance less prominent. The exact change depends on the smoothing algorithm and the signal.

That is the trap worth naming. A ringing resonance, the boxy hum in a room mic, the note a snare shell will not let go of, is precisely a narrow, steady spike. Narrow and steady is the one combination that frequency smoothing is built to erase and that time averaging is built to reveal. So the hunting configuration is the reverse of the pretty one: smoothing off or minimal, averaging slow, or a peak-hold trace. Broadband content averages toward its mean and the parked spike accumulates until it stands up out of the floor.

Pre and post answer different questions

An analyzer inside an EQ can tap the signal before your curve or after it. The distinction shapes the whole workflow.

The failure mode is diagnosing from post. Every band you move reshapes the trace you are reading, so the target drifts with every adjustment and you end up chasing your own EQ. Diagnose on pre, where the spike holds still no matter what you do. Then flip to post to confirm the cut is centered on the spike and deep enough, and that the rest of the curve did not tilt something you did not intend.

The resonance hunt, step by step

With the display configured honestly, the hunt itself is short. Listen first and form a guess: boxy suggests low mids, honky suggests upper mids, a ring on a drum is usually a specific note. Then read the pre trace, unsmoothed, slow averaged, and look for a spike that stays parked while the music moves around it. Drop a narrow bell there, solo the band so you hear only what it passes, and tighten the bandwidth until the ring isolates from the material around it. Flip the boost to a cut and take only as much as it needs to stop poking out.

• Notes are peaks too. A spike that moves with the performance may be the music. A fixed resonance often stays at one frequency across different notes and sections. Watch the trace over time before you commit to cutting anything.
• Check the multiples. Rings and room modes rarely come alone. If the fundamental parks at one frequency, look for smaller spikes near two and three times it before declaring the job done.
• Solo exaggerates. Band solo is for finding the frequency, not for choosing the depth. A ring that sounds offensive in isolation may need only a few dB of cut in context. Set the depth with the solo off.
• A boosted sweep biases you. Sweeping a 12 dB narrow boost makes every frequency sound like a problem. If you sweep, sweep to locate, then reassess at unity before cutting.

A worked example

Here is the loop as it runs in Bevel EQ, which draws its real-time analyzer directly under the summed curve and can draw the pre and post traces at the same time (Options, Show pre and post together).

Say a room mic has a boxy ring. Let it run while the section plays: a spike holds its position through the chord changes, suggesting a fixed resonance. Confirm it by listening. Double-click the curve and engage that band's solo to hear the region the band works on; tighten Q until the ring sits alone. Because frequency, gain and Q are typeable, you can set the center exactly where the spike reads. Flip the gain negative and kill the solo.

Now the two traces do the work for you: the post trace should show the spike reduced while the pre trace still shows it, so you can see the cut and the original in one glance instead of toggling between them.

Save the result as an A/B state and toggle against the untouched state at matched loudness to confirm the cut helps the ring without dulling the room. Close the display for the final listening check.

Keep the reference levels consistent

For an audio spectrum comparison, use the same input level, frequency range, FFT size, tilt and averaging. A trace near the bottom of the display may be the recording's noise floor or the analyzer's chosen range. It is not automatically a frequency that needs boosting. A spectrum's per-frequency levels are also different from overall sample peak or loudness. Use the LUFS and true peak meter for those measurements.

FabFilter's analyzer reference documents how its resolution, speed and tilt settings change the display. Read the equivalent controls in your own analyzer before comparing screenshots from different tools.