What is convolution reverb?

Convolution reverb in plain English

Clap once in an empty stairwell and listen. The clap is over in a few milliseconds, but the stairwell keeps answering for a second or two: a handful of distinct slaps off the nearest walls, then a wash that slowly dies away. Record that answer and you have the stairwell's impulse response, usually shortened to IR.

A convolution reverb loads a recording like that and applies it to whatever you feed in. Your vocal comes out sounding as if it had been sung in the stairwell, because the plug-in has worked out what the stairwell would have done to every moment of it.

An algorithmic reverb works the other way round. It builds a space out of delays and filters, so its size and shimmer can change as you turn the knobs, and the algorithmic vs convolution reverb comparison covers how to choose between them. This page stays with convolution: what an IR is, what convolution does with it, how IRs get made and what the controls on a convolution plug-in change. The reverb lesson covers the controls every reverb shares.

What is an impulse response?

An impulse is the shortest sound there is. In digital audio it's one sample at full level with silence on both sides, and it contains every frequency at once. An impulse response is what comes out when you put an impulse into something, whether that's a concert hall, a hardware plate reverb or a guitar speaker with a microphone in front of it.

A recorded room IR has its parts in a fixed order. First there's a short silence while the sound travels from the speaker to the microphone, about 3 ms for every metre between them. Then the direct sound arrives, followed by the early reflections off the nearest surfaces, and finally the tail, where reflections pile up too thickly to hear apart and fade away. How long that fade takes is the space's RT60, covered in the decay time lesson.

IRs are ordinary audio files, mostly WAV or AIFF. A mono IR has one channel and a stereo IR has two, recorded with a pair of microphones. A true stereo IR has four: the space is recorded twice, once with the speaker on the left and once on the right, with both microphones running each time, so the reverb knows which side a sound came from.

In the lab, the Click card is a perfect impulse. Play it at 100% mix and the vocal comes out untouched, and the panel beside the IR screen measures the difference from the dry loop as silence.

Reading the lab

Each card at the top loads a different impulse response. The Room, Hall and Plate cards are spaces the lab generates in code with the same routine as the main reverb lab, since recordings of real buildings come with licences. Click and Echoes are plain clicks, and Speaker is a 30 ms IR shaped like a guitar cabinet. The Load an IR file button takes any IR you have, and the file never leaves your device.

The big screen shows the whole loop, with the dry sound pale and the convolved sound in colour. Under it, the impulse response screen draws the IR the convolver has right now. Height there is in decibels, from the loudest moment at the edges down to 60 dB quieter on the centre line, so the quiet end of a tail still shows up. The pale shape behind is the IR as it was loaded, which lets you see what trimming, stretching and reversing have done to it.

The panel on the right measures the IR and the loop and says what your settings are doing. Press A/B to hear the dry sound: the lab matches its loudness to the processed version when you do, so the space is the only thing that changes. You can also swap the vocal for another loop, upload a sound or record one.

How convolution works

Digital audio is a stream of samples, 44,100 a second at CD quality, and you can treat each one as a tiny click with its own height. Convolution takes every one of those clicks, sets off a copy of the impulse response scaled to its height, and adds all the copies together. A sample at half level starts a copy at half level; a negative sample starts an upside-down copy.

Put two claps 250 ms apart through a hall IR and you hear two reverbs, overlapping. Put a vocal through it and you get tens of thousands of overlapping copies every second, and their sum is the vocal in the hall. All the plug-in does is add up scaled copies of a recording, which is why a convolution reverb sounds so exactly like the space that was recorded.

The Echoes card shows the same thing from the other side. Its IR is four clicks, at 0, 250, 500 and 750 ms, so convolving with it gives four copies of the sound, which is a delay with three repeats. Any effect that does exactly the same thing every time and doesn't distort can be captured as an IR this way, and that includes EQs and filters as well as rooms.

Swap the echoes for a room and the copies sit so close together that they blur into one sound. That's all a room IR is, as far as convolution is concerned: a very dense set of echoes, each a little quieter and duller than the one before.

Why convolution takes so much processing

Done sample by sample, convolution is a lot of arithmetic. A 2-second IR at 44.1 kHz is 88,200 samples long, and every output sample needs one multiplication for each of them: 88,200 × 44,100 comes to about 3.9 billion multiplications for each second of each channel. Plug-ins get the same answer far more cheaply with the fast Fourier transform (FFT), and most split the IR into blocks so they can start without waiting for the whole thing. A long IR on many tracks still costs more than most algorithmic reverbs, which is one more reason to run a single reverb on a send.

How an impulse response is recorded

The simplest way is to make one short, loud sound in the space and record it. Starter pistols, balloon pops and wooden clappers have all been used. They're quick, but none of them makes a clean impulse: each is louder at some frequencies than others, and they put very little energy into the low end, so the bass of the room gets lost under background noise.

Commercial IR libraries are usually recorded with a sine sweep instead. A speaker plays a tone that glides from deep bass to the top of the hearing range over several seconds, and the microphones record the space's answer. That recording sounds like a sweep smeared by the room. Software then convolves it with the same sweep played backwards, with its level tilted to undo the sweep's balance, and the sweep collapses into a single click with the room's response attached. The process is called deconvolution, and the exponential sweep method most people use was published by the Italian acoustician Angelo Farina in 2000.

A sweep beats a bang because it puts far more energy into the room, spread over its whole length, so the room's answer stands further above the background hum. In the figure the lab measures its own hall the same way, with noise added 50 dB under the sweep, and the measured IR follows the real one until it sinks into the noise. Make the sweep longer and the noise floor drops by about 3 dB for every doubling. The exponential sweep has another advantage: distortion from the speaker ends up just before the IR in time, where it's easy to cut off.

Where the microphones stand changes the IR as much as the room does. Near the stage the direct sound is loud and the reverb is light, while from the back of a hall the reverb takes over, so IR libraries often record several positions in each space.

The same method captures hardware. Send a sweep through a plate, a spring or a vintage digital reverb, record what comes back and deconvolve it, and you have that unit at one setting. Anything that moves inside the unit, such as the slow modulation some digital reverbs use to stop their tails ringing, is lost, because convolution can only reproduce an effect that does the same thing every time.

The controls on a convolution reverb

Logic's Space Designer, Ableton's Hybrid Reverb, Audio Ease Altiverb, Cubase's REVerence, FL Studio's Fruity Convolver and Reaper's ReaVerb all load IRs, and most of them add a similar set of controls for reshaping the IR once it's loaded.

Choosing the IR

The IR decides most of the sound, so pick one close to the space you want before you touch anything else. Trimming and stretching can nudge a hall towards a smaller room, but they can't make a cathedral sound like a vocal booth.

Pre-delay

Pre-delay puts silence in front of the IR. A recorded IR already has a little, from the time the sound took to reach the microphone, so start at zero and add more until the front of each word or hit sits clear of the reverb. On vocals 10 to 30 ms is a common range, and the pre-delay lesson goes into it further.

Start

Start, sometimes labelled offset or trim, cuts the beginning off the IR. On a recording that removes the silent gap and the direct sound, and you usually want the direct sound gone: the dry track already supplies it, and a second copy arriving a few milliseconds later on the reverb return causes comb filtering. Push the start further and it cuts into the early reflections, which leaves a reverb with less sense of the room's shape. In the lab, turning it up on the hall shows exactly how much of the IR's energy goes with each millisecond.

Length and envelope

Length uses less of the IR, measured from the start point, with a short fade at the end so it doesn't click. Many convolution plug-ins also have an envelope that reshapes the IR's level over time, so a 4-second hall can be made to fade out in 2 seconds without sounding chopped. The lab only has length, which shows what happens when you take too much off: at 20% the hall stops while it's still loud, and that abrupt end is close to the gated reverb sound of 1980s drums.

Stretch

Stretch makes the IR longer or shorter. The lab does it by resampling, which is the same as playing the recording slower or faster, so a stretched IR also comes out darker and a squeezed one brighter. If a plug-in calls the control size, listen for whether the tone changes along with the length.

Reverse

Reverse plays the IR backwards, so each sound is followed by a swell that stops dead. The sound you feed in still plays forwards, which means the swell comes after each word. The reverse reverb effect that rises into a word is made a different way, by reversing the audio itself, and the reverse reverb lesson shows how.

EQ and mix

Most convolution plug-ins have filters on the reverb. A low cut somewhere between 150 and 300 Hz keeps a long IR from thickening the bass, and a gentle high cut pushes the reverb further back. On a send, set the mix to 100% wet and use the send level instead; as an insert on a vocal, start somewhere around 15 to 25% and adjust by ear.

Beyond rooms: speakers and other IRs

An IR doesn't have to be a space. A guitar speaker with a microphone in front of it is a filter, and almost all of its impulse response is over within a few tens of milliseconds. Guitar amp simulators load cabinet IRs for that reason, and there are IR loader plug-ins that do nothing else. The lab's Speaker card is a 30 ms IR shaped like one: the lows under about 90 Hz and the highs over about 5 kHz fall away, with a bump near 2.5 kHz. Put the guitar through it at 100% and the convolution changes the tone without adding any space at all.

How long an IR rings decides how you hear it. When it's all over in a few milliseconds your ear can't separate its parts, so it comes across as tone colour, and once parts of it arrive far enough apart to hear on their own it starts to sound like a space or an echo.

You can also load any audio file as an IR. A snare hit, a cymbal or a few seconds of rain gives a space no building could have, since every sample of your track sets off a copy of that file. Results are often messy; short sounds that die away make the easiest starting point. Try it in the lab with the Load an IR file button.

For more on the character of each space, the plate vs hall vs room vs spring comparison plays them side by side.

How to use convolution reverb, step by step

1. Put it on a sendMake a return track with the convolution reverb on it at 100% wet, and send to it from the tracks that need space. One reverb costs less CPU and puts everything in the same room. The insert vs send lesson explains the routing.
2. Choose the IR with the track playingTry a small room for drums and guitars that should feel close, a plate for vocals, and a hall for strings and pads. Audition several in the mix, since an IR that sounds lovely on its own can swamp a busy arrangement.
3. Trim the startIf the IR begins with a gap or a loud direct sound, turn the start up until it begins at the first reflection.
4. Add pre-delayBring it up until the front of each word or hit sits clear of the reverb, often somewhere between 10 and 30 ms on a vocal.
5. Shorten it if the tail runs into the next phraseUse the length or the decay envelope. If you're taking off more than half, a smaller IR will usually sound more natural.
6. EQ the returnCut below 150 to 300 Hz, and roll the top off if the reverb is brighter than the track.
7. Set the level, then compareUse the send level, and check with the reverb muted at matched loudness so you hear the space and nothing else.

Common convolution reverb mistakes

Leaving the direct sound in the IR

Many recorded IRs start with the direct sound, a loud first click. On a send that adds a second, slightly late copy of the dry track, which thins out the sound or makes it ring. Turn the start control up until that first click has gone.

Choosing an IR by its name

A file called Vocal Plate won't suit every vocal. IRs are recordings of particular spaces with particular microphones, so judge them by ear on your track and expect to try several.

A long IR on every track

Ten instances of a 4-second hall cost a lot of CPU, and once each one has been trimmed and filtered a little differently, the parts no longer sound as if they were in the same room. One send does the job better.

Expecting the tail to move

Convolution plays back the same IR every time, so a long tail behind sustained pads can sound frozen next to an algorithmic reverb with modulation. If it bothers you, try an algorithmic reverb on those parts and keep the convolution for drums and vocals.

Comparing at different volumes

Adding reverb usually makes a track louder, and louder tends to sound better in a quick comparison. Use the A/B in the lab, or match the levels by ear in your DAW, before you decide an IR is helping.

Train your ears

The quiz under the lab loads a hidden IR at matched loudness. Some rounds ask whether the hall is playing forwards or reversed, and others ask whether you're hearing the room, the hall or the plate. Eight rounds take a few minutes, and your best score stays in your browser.

Questions people ask

Keep going