Reach for an algorithmic reverb by default: every control works while the music plays, and its modulated tails suit vocals, pads and anything with long notes. It's lighter on your computer, too.
Use a convolution reverb when you want one specific space or box, such as a real studio room around acoustic drums or a sampled vintage plate on a vocal. You get that sound exactly, and you give up most of the freedom to change it.
At a glance
| Property | Algorithmic | Convolution |
|---|---|---|
| How it works | Recirculates the sound through delay lines, filters and gains | Multiplies the sound against a stored impulse response |
| Where the space comes from | The maths and its settings | A recording of a room, a plate, a spring or a hardware unit |
| Changing the space | Every control reshapes it as it plays | Edit the recording: trim, fade, stretch, filter |
| Movement in the tail | Can be modulated so it never repeats | The same recording every time |
| CPU and memory | Low; well under a second of delay lines | Higher, and it grows with the IR's length |
| Latency | Normally none | None to a few milliseconds, depending on the plug-in |
| Best at | Vocals, pads, huge or impossible spaces, sound design | Real rooms, halls and sampled hardware |
| Examples | Valhalla VintageVerb, FabFilter Pro-R, Logic's ChromaVerb, Ableton's Reverb | Audio Ease Altiverb, Logic's Space Designer, Steinberg REVerence, FL Studio's Fruity Convolver |
What is algorithmic reverb?
An algorithmic reverb is a set of delays that feed back into themselves. A delay line holds the sound for a few tens of milliseconds and lets it out again. Feed its output back to its input and you get a train of echoes; connect several delay lines of different lengths so each one feeds all the others, and the echoes multiply. In the lab's network every trip round sends each line's output into all eight, so one click can become up to 8 echoes, then 64, then 512, within about the first 300 milliseconds. By then they're far too close together to hear as separate repeats, and they've become a reverb tail.
Manfred Schroeder published the first well-known design at Bell Labs in 1962, using a handful of feedback delays and allpass filters. Studios got digital reverb in hardware with the EMT 250 in 1976 and the Lexicon 224 in 1978. Many modern algorithms descend from the feedback delay network that Jean-Marc Jot described in the early 1990s, and that's what the lab runs: four allpass diffusers that smear each sound into a burst, then eight delay lines mixed together on every trip.
Each trip round costs the sound a little level, and that loss sets the decay. With the lab's decay at 1.8 seconds and size at 50%, the longest delay line is 85 ms, so it loses 60 × 0.085 ÷ 1.8 = 2.83 dB per trip. About 21 trips take it down 60 dB, and that time is the RT60 the decay knob sets. A low-pass filter in every loop takes more off the top end on each trip, which is damping, so the lab measures the whole tail at 1.58 seconds, a little under the 1.8 on the knob. Change any setting and the lab recalculates those losses and delay times on the spot, which is why an algorithm can follow a knob as you turn it.
What is convolution reverb?
A convolution reverb starts from an impulse response (IR), a recording of how a space answers one instant click. To make one, an engineer plays a sine sweep through a speaker in the room and records the result, then software turns that recording into the room's answer to a click. Convolution sets off a copy of the IR for every sample of your track, scaled to that sample's level, and adds the copies up. The convolution reverb lesson covers the maths and the recording process in detail.
Sony brought it to studios in hardware with the DRE-S777 in 1999, and Audio Ease's Altiverb made it a plug-in in 2001. The big change was where the space comes from. A convolution reverb has no model of a room at all; it plays back whatever the IR contains, which might be a cathedral, a cupboard, an EMT plate or a Lexicon on one particular preset.
Can a convolution reverb copy an algorithmic one?
Sometimes perfectly, and the second lab step shows it. With modulation off, the lab's algorithm does exactly the same thing to every sound, at any moment and at any level. Engineers call a process like that linear and time-invariant, and its impulse response describes it completely. So the lab plays one click through the algorithm and loads the result into a convolver, and the convolver makes the same reverb. The Right now panel under the screens measures the difference between the two on your loop: with modulation off it's about 80 dB under the reverb, far too quiet to hear.
Movement doesn't survive the trip. An algorithm with modulation changes from moment to moment, so a capture holds one instant of it. Anything that depends on level is lost as well, such as a reverb that saturates when you hit it hard or one that ducks its own tail while you play. And one IR is one setting, so sampling a hardware unit properly means recording an IR for each preset and decay time you want to keep, which is why sampled hardware libraries come with dozens of them.
Modulation: why algorithmic tails can sound more alive
Many algorithmic reverbs let their delay lines drift slowly in length, moved by an LFO. In the lab each line can swing by up to 0.8 ms, at its own speed between 0.31 and 1.27 times a second, which bends the pitch of the tail by about 11 cents at most. That's a small change. Even so, the tail never meets the same set of echoes twice, and that smooths out the metallic ringing a long, still tail can have. Lexicon's Random Hall calls its version of this spin and wander.
The two-clicks screen in the lab measures this, by sending one click into the reverb and another half a second later and comparing the answers. With modulation at 70% they differ almost completely, sample by sample, even though they sound like the same space. Through the convolver they're identical, because the IR is the same recording every time.
You'll hear the difference most on held notes: strings, pads, organ and sustained vocals, where a long convolution tail can sound frozen next to a modulated algorithm. On drums and short rooms it's much harder to hear, because the tail is gone before there's time for it to move.
Reshaping a space vs editing a recording
The controls on an algorithmic reverb change the network itself. Size changes the lengths of the delay lines and decay changes how much they lose on each trip; damping, diffusion, modulation and the early reflections each work on their own part of the structure. You can make a 30 metre hall with the damping of a tiled bathroom, or a 20 second tail that never quite ends.
A convolution reverb's controls edit the recording. You can delay it with pre-delay, cut the start off, shorten it with a fade, filter it, or stretch it like a sample played slower, which also makes it darker. Many plug-ins put a volume envelope on the IR, and Logic's Space Designer lets you draw one. That covers a lot, but every one of those controls works on what's already in the file. A fade can make a 2 second hall last 1 second. Making it last 4 takes a stretch, and stretching changes its tone along with its length.
CPU, memory and latency
The lab shows the memory side as Samples held. At its starting settings the algorithm keeps about 22,000 samples in its delay lines and diffusers. The IR it captures from those settings is 2.4 seconds of stereo, about 212,000 samples and so roughly ten times as many, and a 6 second stereo hall IR at 48 kHz holds 576,000.
Processing follows the same pattern. The lab's algorithm does roughly a hundred multiplications per sample, whatever its decay. Convolution done directly needs one multiplication per IR sample for every output sample, so that 2.4 second IR would need 105,840 for each sample of each channel. No plug-in does it that way: they use the fast Fourier transform, which brings the cost down by a very large factor, and split the IR into blocks. Some of those designs add a few milliseconds of latency, which your DAW's delay compensation hides. A long IR still usually costs more than an algorithm, and the cost grows with its length, so convolution belongs on a send shared by several tracks.
Which should you use?
| Job | Algorithmic | Convolution |
|---|---|---|
| Lead vocal | Good: a modulated plate or hall | Good with a sampled plate |
| Pads, strings and long notes | Best; the tail moves | Can sound frozen on long tails |
| A real room around acoustic drums | Works | Best; it's the actual room |
| Orchestral or acoustic realism | Works with care | Best, with a good hall IR |
| The sound of a vintage unit | Use an emulation of it | Good for its tone, missing its movement |
| Huge or impossible spaces | Best | Only if someone made the IR |
| Lots of instances on a laptop | Best | Use one on a send |
For most pop and electronic mixes, start with an algorithmic reverb. The vocals and synths that carry those songs gain the most from a tail that moves and a decay you can set to the tempo, and you can run more instances of it. Switch to convolution when a part needs to sound like it was played in a particular place, which is common with acoustic drums, piano, strings and anything recorded with room microphones you want to extend.
Hybrid reverbs
You can also use both at once. A common pairing is a convolution reverb for the first hundred milliseconds or so, where the early reflections of a real room tell your ears most of what they learn about where you are, and an algorithmic tail after it that you can lengthen and modulate. Ableton Live's Hybrid Reverb puts both engines in one device. In any other DAW, put a short room IR and an algorithmic hall on the same return track, or on two returns, and balance them.
How to choose between them, step by step
- Decide whether you need a particular placeIf the brief is a real room, a known hall or a sampled piece of hardware, go convolution. If any good space will do, go algorithmic.
- Listen to the longest notesSolo the reverb return on a held chord or a sustained word. If the convolution tail sounds static or ringing, try a modulated algorithm on that part.
- Set the decay to the songOn an algorithm, turn the decay until the tail clears before the next phrase. On a convolution reverb, pick an IR near the length you need and trim it with the length or envelope controls, so you don't have to stretch it far.
- Put it on a sendOne reverb of either kind at 100% wet on a return track, shared by the parts that belong in that space.
- Compare at matched loudnessMatch the levels of the two candidates before you choose, because the louder reverb will seem better for no other reason.
Common mistakes
Expecting a convolution decay knob to act like an algorithm's
On most convolution plug-ins, decay is a fade on the recording or a stretch of it. Fades can only shorten, and big stretches change the colour of the tail along with its length. If you need a much longer tail, load a longer IR.
Assuming convolution is always the realistic choice
An IR is a recording, and it carries the microphones, the speaker and the noise of the room where it was made. A clean algorithmic hall can sit better behind a vocal than a noisy IR of a famous church, so judge by ear with the rest of the mix playing.
One convolution reverb per track
Ten instances of a 4 second IR will load your computer heavily, and once their settings drift apart the parts stop sounding as if they share a room. Use one or two on sends.
Sampling a modulated reverb and wondering where the shimmer went
A capture holds one moment of a moving tail, as the sixth lab step shows. If the movement was the reason you liked that unit, use an algorithmic emulation of it, or put a slow chorus on the convolution return to add some back.
Training your ears to tell them apart
The quiz under the lab gives both engines the same modulated hall at the same loudness and asks which one is playing, so the only clue left is whether the tail moves. Pick the strings or chords before you start and listen in the gaps after each chord, where the tail is on its own. Some rounds ask instead whether the convolver's IR has been cut short.
Questions people ask
Is convolution reverb more realistic than algorithmic?
A convolution reverb loaded with a good recording of a real room is that room, so for realism it's very hard to beat. A well-made algorithm can sound just as convincing, though, and a convolution reverb is only as real as its IR: a noisy or badly placed recording sounds like exactly that.
Why do people still use algorithmic reverb?
Because you can change everything about it while it plays, including spaces that couldn't exist, and its tail can move. It also costs less CPU than a long IR. The Lexicon 480L and the Bricasti M7, two of the best-known studio reverbs, are both algorithmic.
Can I make an impulse response of my hardware reverb?
Yes. Play a sine sweep through the unit, record the output and deconvolve it, the same way rooms are measured. You'll capture the tone and decay of that one setting. Any modulation, and any change in how the unit behaves at different levels, won't come across.
Which uses more CPU?
Usually the convolution reverb, and the longer the IR the bigger the gap. The lab's algorithm does roughly a hundred multiplications per sample, while a 2.4 second IR done directly would need 105,840 per sample per channel. Plug-ins use FFT methods to cut that down a long way, but a long IR is still the heavier of the two.
Is a hybrid reverb worth it?
It's a good way to get a real room's early reflections with an adjustable tail behind them. Ableton Live's Hybrid Reverb puts a convolution engine and an algorithmic one in a single device, in series or in parallel, and you can do the same in any DAW by putting both on one return track.
Does it matter which type I put on a send?
Both work the same way on a send at 100% wet, so choose by the sound you want and how many instances your computer can run. One long convolution reverb shared by several tracks on a send costs far less than one on each track.