Three EQ moves that cover most voices

Cut first, boost rarely, and know which frequencies you are touching

Most voice EQ is three moves. They are the same three moves nearly every time, and the reason they work is that they address three specific and predictable problems rather than a taste.

1. High-pass filter. Remove what is below the voice.

Adult speech fundamentals sit roughly between 85 and 180Hz for typical male voices and 165 to 255Hz for typical female voices. Below the fundamental there is nothing but rumble: traffic, air handling, footsteps, desk knocks, handling noise, and the low-frequency half of plosives.

Set a high-pass at 80Hz for a low voice, 100 to 120Hz for a higher one, with a slope of 12 or 18dB per octave. Sweep it upward while listening until the voice starts to thin, then back off. This is the single most useful filter in speech work and it is doing nothing to the voice at all: it is removing energy that was never part of it and that costs you headroom in every later stage.

2. Cut the mud. There is almost always a build-up between 200 and 400Hz, from proximity effect, from a desk or a boundary reflection, or from the room. A 2 to 4dB cut with a moderately wide Q is usually right. Too much and the voice becomes thin and telephone-like.

3. A small presence lift. Consonant clarity lives between about 2 and 6kHz. A gentle 1 to 3dB broad boost in that region improves intelligibility noticeably. Above 10kHz, a shelf adds “air”: worth a decibel or two on a good recording and worth nothing on a noisy one, because up there you are mostly lifting hiss.

That is the whole default. Most voices need nothing else.

Finding a resonance rather than guessing

When something specific is wrong (a honk, a boxiness, a ring) there is a reliable method.

Take a narrow band, boost it by 10 to 12dB, and sweep slowly across the spectrum while the voice plays. At the offending frequency, the problem becomes unbearable. Note the frequency, then cut there instead, by 3 to 6dB, with a moderate Q.

This works because a resonance is narrow and amplifying it makes it unmistakable. It is also easy to overuse: the sweep makes everything sound bad at high boost, and there is always a worst frequency. If nothing jumps out as obviously wrong, there is nothing to fix.

Cut narrow, boost wide

A useful asymmetry. Problems tend to be narrow (a resonance at one frequency, a ring from one room dimension) so cuts can be surgical. What sounds pleasant tends to be broad, because tonal character is a shape across a region rather than a spike. A narrow boost sounds like a resonance, because it is one.

Practical: cuts at Q of 2 to 6; boosts at Q of 0.7 to 1.4.

Minimum phase, linear phase, and what each costs

An ordinary EQ is minimum phase: changing the amplitude at a frequency necessarily shifts the phase around it. This is not a defect but a mathematical consequence of a causal filter, and it is what analogue EQ does. On a single voice it is inaudible and unimportant.

A linear phase EQ keeps phase constant by processing with a symmetric filter, which requires looking ahead, so it introduces latency, and it produces pre-ringing: a faint smear before a transient, which has no physical analogue and can sound unnatural on percussive material.

Which to use: minimum phase for almost everything, and particularly for anything you are monitoring live. Linear phase is worth it when you are EQing two microphones on the same source that will be summed, because a phase shift on one and not the other changes how they combine. On a single voice track, reach for the ordinary one.

The point everybody skips

Fix it with placement before you fix it with EQ. A boxy recording caused by a desk reflection has a notch in it, and an EQ boost aimed at a notch is amplifying the noise in a band where the signal was cancelled. Moving the microphone 20cm or putting a cloth on the desk removes the cause.

EQ is very good at removing excess energy and very bad at restoring energy that cancelled. Cuts are cheap; boosts have a cost.

A worked starting point

For a close-miked voice in a treated-enough room:

high-pass       90 Hz,  18 dB/oct
bell           280 Hz,  −3 dB,  Q 1.2
bell          3.5 kHz,  +2 dB,  Q 1.0
high shelf     11 kHz,  +1.5 dB

Then listen and change one thing. The mistake is to build a six-band chain because the plugin has six bands.

The free path

Audacity has a graphic and a parametric EQ and a spectrogram view. Ardour ships with a capable parametric EQ and analyser. The Calf and LSP plugin suites are free, cross-platform, and professional in scope. ReaEQ, from Reaper, is free to use within Reaper’s evaluation. ffmpeg applies EQ from the command line for batch work:

ffmpeg -i in.wav -af "highpass=f=90,equalizer=f=280:t=q:w=1.2:g=-3,equalizer=f=3500:t=q:w=1:g=2" out.wav

A spectrum analyser costs nothing and is worth more than any EQ: Voxengo SPAN is free, and Ardour and Audacity both include one.

The limitation

EQ changes the balance of what was recorded and cannot separate two things occupying the same band. If the fridge hum and the voice’s fundamental are both at 120Hz, a cut at 120Hz removes both. Every EQ decision on a noisy recording is a trade rather than a repair.

It also cannot fix a room. A comb filter from a reflection has notches every few hundred hertz, and EQ can only apply a smooth curve to a jagged problem. You can reduce the most prominent tooth and you cannot fill in the gaps, because the cancelled energy is not attenuated: it is gone.