What is comb filtering?

Comb filtering in plain English

Take any sound, make a copy, delay the copy by a millisecond and add the two back together. You might expect a slightly louder version of the same sound. What you get has a strange, hollow tone, because the delay has knocked some frequencies out completely and made others louder.

That happens because a 1 ms delay means something different to every frequency. A 500 Hz wave takes 2 ms to go through one cycle, so a 1 ms delay leaves the copy exactly half a cycle behind: every push from the original meets a pull from the copy and they cancel. A 1 kHz wave takes 1 ms per cycle, so the copy is one whole cycle behind and lines up again, and the two add up to twice the level. Every frequency in between lands somewhere between those two extremes.

Draw the result as a frequency response and you see a row of deep, narrow dips at evenly spaced frequencies with rounded peaks between them, like the teeth of a comb. That's the comb filter. The phase lesson covers phase and polarity in general; this page stays on the comb itself, from where its notches land to how you get rid of one you didn't ask for.

Where the notches land

Everything about the pattern comes from one number, the delay between the two copies. The first notch is at 1 ÷ (2 × delay). After that, a notch arrives every 1 ÷ delay, so they sit at odd multiples of the first one: three times, five times, seven times and on up the spectrum.

With a 1 ms delay that's 1 ÷ 0.002 s, so the first notch is at 500 Hz and the rest are at 1.5 kHz, 2.5 kHz, 3.5 kHz and every 1 kHz after that. The peaks sit halfway between, at whole multiples of 1 kHz, where the copy is a whole number of cycles late. Double the delay to 2 ms and every notch halves in frequency: the first moves down to 250 Hz and the rest follow it, now only 500 Hz apart.

So a longer delay pulls the comb down the spectrum and packs the teeth closer together, and a shorter one pushes it up and spreads them out. Drag the delay in the figure and watch both happen at once.

The distance column is there because sound travels about 343 metres a second at room temperature, which is roughly 34 cm every millisecond. Any time a sound reaches the same place by two paths of different length, the difference in length sets the delay, and the delay sets the comb.

How deep the notches go

The delay decides where the notches are, and the level of the copy decides how deep they go. When the copy is as loud as the original, the two cancel perfectly at each notch frequency and nothing is left there at all, while the peaks come out 6 dB louder than the sound on its own. Turn the copy down and the cancellation is only partial, so the notches get shallower and the peaks get lower.

The gap between a peak and a notch is 20 × log10((1 + g) ÷ (1 − g)) dB, where g is the copy's level as a fraction of the original. You won't need to work it out by hand, but these numbers are useful to know.

The 33% row is the reason for the 3:1 rule for placing microphones, which says to keep each mic at least three times as far from any other mic as it is from its own source. In open space a sound three times further away arrives at about a third of the level, roughly 9.5 dB down, so the spill between the mics makes a comb with only 6 dB between peak and notch. You can still hear it if you listen for it, but it no longer takes chunks out of the sound.

What does comb filtering sound like?

On a single steady sound it's hard to point at, which is why the lab starts with pink noise, a hiss with energy spread across the whole range. Add a 1 ms copy to it and the hiss turns hollow and nasal, like listening through a cardboard tube. People often describe a comb-filtered recording as phasey or boxy.

The character changes with the delay. Under about 0.2 ms the first notch is up at 2.5 kHz or higher and only a few teeth fit in the range you can hear, so the sound mostly loses some top end and comes across as duller. From about 0.5 ms to 3 ms the notches fall right across the midrange, where most of a voice or an instrument's tone lives, and you get the strongest hollow sound.

Between about 3 ms and 20 ms the teeth are packed so close together that you stop hearing single dips. The peaks are now evenly spaced like the harmonics of a note, at whole-number multiples of 1 ÷ delay, so the sound picks up a faint metallic pitch at that frequency: a 5 ms delay gives a buzz around 200 Hz, and 10 ms one around 100 Hz.

Past roughly 30 ms the two copies stop blending into one tone. The ear starts to hear two events, first a doubled sound and then an echo, and although the comb pattern is still there on paper, it stops being what you notice. The Haas effect lesson covers the region where that changes over.

Reading the lab

The lab can make a comb filter in two ways. Copy level and copy delay add a second copy of the sound inside each channel, the way a second microphone would. Delay R plays the whole right channel late, which only turns into a comb filter when the two channels are added together, and the Mono button does that adding. Width and pan belong to the stereo width lesson, so they're locked here.

The wide screen is the one to watch. The lab sends a click through your settings and measures the frequency response, with low frequencies on the left and high on the right. The pink line is what you'd hear in mono and the dashed line is the left speaker on its own. When the comb comes from the copy, both lines show it, because each speaker plays it; when it comes from Delay R, only the pink line dips, because that comb only exists once the sides are added. Each notch is labelled with its frequency in Hz, and you can hover over the screen to read any point.

The goniometer and the SUM meter show how much the two channels have in common, and the two-cycle screen sends a 110 Hz tone through your settings so you can see the right side slide along when you delay it. The Right now panel explains what the current settings are doing, with the notch frequencies and the mono loss measured from the loop. You can also upload your own sound, or record eight seconds of one, and it plays through the lab without leaving your device.

Where comb filtering comes from

Two microphones on one source

This is the textbook case. Put one mic close to a snare and another a metre further back, and the second one hears each hit about 2.9 ms later. Mix them and the first notch lands near 172 Hz, right where the snare's body is, with more notches every 345 Hz above. You'll meet it most on multi-miked drums and on guitar cabs with two mics.

The further mic is usually quieter than the close one, which is why multi-miked drums sound coloured far more often than they sound broken. Nudging the later track earlier until the hits line up removes the delay, and the comb goes with it. The drum phase alignment lesson works through a whole kit.

Reflections near the microphone

One mic can comb filter on its own, because it can hear the source twice: once directly and once bounced off a nearby surface. Say a singer stands at a desk with the mic between them and the desktop, and the bounce travels 34 cm further than the direct sound. That's a 1 ms copy, and it puts the first notch at 500 Hz. Hard, flat surfaces close to the mic do the most damage, since they reflect the most and the reflection arrives barely quieter than the direct sound. These early reflections are a big part of why a vocal recorded in a small, bare room can sound boxy on any mic.

Your monitors have the same problem. Sound from each speaker reaches your ears directly and also off the desk in front of you, so a big, flat desk surface between the speakers and your head can put a comb on everything you hear while you mix.

Copied parts and layered samples

Duplicating a track and nudging the copy a few milliseconds is a common way to fatten a part, and it's a comb filter by definition. Layering two drum samples that share a recording, or that both start with a similar click, does the same if their starts are slightly offset. A real second performance avoids this, because the timing between two takes keeps shifting, so no fixed notch pattern builds up.

Latency in parallel chains

Any time a signal is split and recombined, both paths need to take exactly the same time. A plug-in that adds 64 samples of latency at 48 kHz holds its path back by 1.33 ms, which would put a notch at 375 Hz and another every 750 Hz when the two paths meet again. DAWs compensate for plug-in latency automatically, so this mostly bites when compensation is switched off or when one path runs through outboard hardware. A plug-in that reports its latency wrongly can do it too.

Even one sample matters at the top. Two equal copies one sample apart at 44.1 kHz lose 3 dB at 11 kHz and cancel completely at 22.05 kHz, which you'd hear as a slightly dull parallel blend.

Stereo tricks summed to mono

A delay between the left and right channels sounds fine in stereo, since each speaker plays a whole sound and nothing gets added electrically. A Haas widener on a guitar works exactly like that. Press mono and the two sides are added, the late side becomes the copy, and the comb filter appears. Stereo wideners and chorus effects built on short delays can do the same. This is the case the lab's Delay R knob shows, and it's the main reason mono compatibility checks exist.

How to fix comb filtering

Comb filtering comes from timing, so the fixes work on timing and level, with EQ last if at all.

1. Find the pairSolo the tracks that carry the same sound, two at a time, and listen in mono. When the hollow tone appears, you've found the two copies.
2. Line them upZoom in until you can see the start of a hit on both tracks and slide the later one earlier until the waveforms match. Logic has a Sample Delay plug-in and Ableton Live has a Track Delay on every track, both for moving a whole track by milliseconds or samples.
3. Then try the polarity switchOnce they're lined up, flip one track's polarity and keep whichever setting sounds fuller. Do it in this order, because flipping first only moves the notches.
4. Turn one copy downIf the two can't be lined up, as with a room mic that's meant to sound distant, keep it at least 10 dB below the close mic and the notches will be shallow.
5. Fix it at the source next timeFollow the 3:1 rule between mics and keep each one well away from desks and walls. A duvet hung behind the singer cuts down how much of their voice the wall behind them bounces back into the mic.
6. Check the routingFor parallel chains, make sure latency compensation is on and both paths reach the mix at the same moment.

EQ comes last because a boost wide enough to cover a notch also lifts the peaks around it, so the dip is still there, only louder.

Using comb filtering on purpose

The same thing that ruins a drum recording is the basis of several effects. A flanger is a comb filter whose delay sweeps slowly up and down, usually somewhere under 10 ms, so the notches glide through the spectrum and give the familiar jet-plane whoosh. The name comes from tape: engineers played the same recording on two tape machines in sync and slowed one down by pressing on the flange of its reel.

Feed some of a comb filter's output back into its input and the notches give way to sharp, ringing peaks at every multiple of 1 ÷ delay. That's a feedback comb filter, and with enough feedback it rings at a clear pitch. Kevin Karplus and Alex Strong published a way to make plucked-string sounds from this in 1983: fill a short feedback delay with a burst of noise and let it ring, and the length of the delay sets the note.

A phaser sounds related and works differently. It uses all-pass filters in place of a delay, which gives a handful of notches that aren't evenly spaced, so it sounds smoother and less metallic than a flanger. The chorus, flanger and phaser comparison puts the three side by side.

Common comb filtering mistakes

Reaching for the polarity switch first

Flipping polarity only helps when two signals are close to opposite at every frequency, as with a snare's top and bottom mics. When the cause is a delay, flipping one copy moves every notch to where the peaks were and adds one at 0 Hz, so a 1 ms delay goes from notches at 500 Hz, 1.5 kHz and up to notches at 0 Hz, 1 kHz, 2 kHz and up. Line the timing up first.

Chasing the notches with EQ

If you find yourself boosting 400 Hz, then 1.2 kHz, then 2 kHz on a track that still sounds hollow, stop and look for the second copy.

Moving one mic of a pair by eye

A few centimetres of extra distance moves the upper notches by hundreds of hertz. Move the mic while listening to both mics together in mono, and stop where the low end is fullest.

Checking only on headphones

Headphones keep the two channels apart, so a Haas widener or a short stereo delay sounds clean on them and still falls apart on a phone speaker. Press mono before you decide a stereo trick is safe.

Train your ears to hear comb filtering

The quiz under the lab plays a loop with hidden settings and asks whether it will survive being played in mono. In the rounds that fail, the right side is a millisecond or so late or flipped; in the others only the width changes, which mono ignores. The screens are covered and the rounds are matched in loudness, so you have to hear the stereo version and predict whether summing it will dig out a comb.

Before you start, spend a minute with the pink noise and the copy delay knob, sweeping it slowly from 0.5 ms to 5 ms. Once you've heard the notches move on noise, the same hollow tone gets much easier to pick out on drums and voices.

Questions people ask

Keep going