Skip to content

Convolution reverb copies a space

A convolution reverb plays your sound through a recording of how a room answers a single click. It captures the space as one fixed snapshot, with no movement or drive.

Updated · 7 min read Lesson 11 of 14 in Audio Science

In short

  1. An impulse response records how a space answers one click: direct sound, early reflections, then a decaying tail.
  2. Convolution starts a scaled copy of that response for every input sample and adds the copies together.
  3. It captures one fixed, linear snapshot: great for real rooms and cabinets, with no modulation and no drive.
In this lesson (8)

A dry vocal looking for a room

You have a clean, dry vocal from a treated booth. It sits on top of the beat instead of inside it. You add a reverb and raise the send, and now it sounds like a reverb: a smooth tail behind the voice, but no sense of a real room around it.

There are two ways to make a reverb. An algorithmic reverb builds a tail from networks of delays and feedback, which makes it flexible and easy to animate. A reverb takes the opposite route. It starts from a recording of how one real space answers a single click, and plays your sound through that recording.

Why it matters

The room is in the first moments

When you clap in a room, the direct sound reaches you first. Then come the , separate echoes from the nearest walls, floor and ceiling, arriving over the first few tens of milliseconds. Sound travels about 34 cm per millisecond, so a reflection whose path is 3.4 m longer than the direct path arrives 10 ms later. After that the reflections pile up into a dense tail that dies away. The early part tells your ear a lot about the size of the space and how far away the source is. The tail sets how long the room hangs on.

Figure 1. A room answering one click, drawn as a shape rather than a measurement: the direct sound, a few separate early reflections, then a dense tail that dies away.

That recording of a room's answer to a click is its . Because it holds the exact pattern of reflections of one real space, a convolution reverb can place a dry recording in that space without any tuning. It works the same way for anything else you have an impulse response for: a plate, a spring, a speaker cabinet, a stairwell you recorded yourself.

Listen

Push a sound forward or back with reverb

A short melody in a generated room. Pre-delay, decay and level each change how far away the notes feel.

Loading demo…
100%

Start with your speakers or headphones low.

The demo above is itself a convolution reverb. It convolves the melody with a generated impulse response of decaying noise, and the decay control sets how long that response is.

The science

Convolution adds up copies of the response

A room behaves, to a good approximation, like a linear time-invariant system: double the input and the output doubles, play the same input later and the same output comes later. For such a system, the response to one impulse tells you the response to anything, because any signal is a series of impulses of different sizes, one per sample. Each input sample starts its own copy of the impulse response, scaled by that sample's value, and the output is the sum of all the copies:

y[n]=∑k=0nx[k] h[n−k]y[n] = \sum_{k=0}^{n} x[k]\, h[n - k]

Here xx is the dry input, hh the impulse response and yy the wet output. Convolution is not the signal multiplied by the response sample by sample. It is this sum of shifted, scaled copies.

Figure 2. Convolution in pictures. Each hit in the input starts its own copy of the impulse response, scaled by how loud the hit was, and the output is all of those copies added together.

Done directly, the cost is large. A 2-second impulse response at 48 kHz has 96,000 samples, so every output sample needs 96,000 multiplications, about 4.6 billion a second for one channel. Plugins use the instead, because convolution in time equals multiplication in frequency. Splitting the response into blocks of growing size keeps that efficient without adding delay between input and output (Gardner, 1995).

To capture a response, engineers play an exponential sine sweep through a speaker in the space, record it and process the recording back into an impulse response (Farina, 2000). A balloon pop or a starting pistol works too, with more noise.

The model has limits. An impulse response captures one source position, one microphone position and one frozen moment. It cannot capture distortion or anything that moves. A convolution reverb has no modulation, and an impulse response of a guitar amp or an analog copies only its frequency and response at one setting, not its drive or the way it changes with level.

Try it

Turn any reverb into an impulse response

You need an algorithmic reverb and a convolution reverb. If your DAW has no convolution reverb, any free one that loads impulse responses works.

0 of 7 steps done

The convolution reverb reproduces whatever made its impulse response, including your own algorithmic reverb, but only as a frozen snapshot.

Common mistake

The biggest hall on a busy mix

A cathedral or a large hall sounds lush in solo, so you choose one. A tail several seconds long fills every gap in a dense arrangement, covers the of the drums and pushes the vocal back. Choose the space for the tempo and density of the song: short rooms for fast, busy tracks, long halls for sparse ones. Put a around 150 Hz on the reverb return so the tail stays out of the kick and bass.

Takeaway

Choose the space for the scene

Pick an impulse response for where you want the listener to be, not for the name of the preset. Sending parts recorded in different rooms to the same room response puts them in one shared space. Use convolution for realism and fixed character, such as rooms, plates and cabinets. Use an algorithmic reverb when you want movement, a tail you can reshape freely or a lighter CPU load.

3 questions

Check yourself

  1. What does a convolution reverb do with each sample of your dry vocal?
  2. Why can an impulse response of a guitar amp not capture its drive?
  3. A 2-second impulse response runs at 48 kHz. Done directly, how many multiplications does each output sample need?

Sources

Show all 4
  • Farina, A. (2000). Simultaneous measurement of impulse response and distortion with a swept-sine technique. 108th AES Convention, Paris, paper 5093.
  • Gardner, W. G. (1995). Efficient convolution without input-output delay. Journal of the Audio Engineering Society, 43(3), 127-136.
  • Smith, J. O. Spectral Audio Signal Processing. CCRMA, Stanford University. ccrma.stanford.edu/~jos/sasp/
  • Smith, J. O. (2007). Introduction to Digital Filters with Audio Applications. CCRMA, Stanford University. ccrma.stanford.edu/~jos/filters/

Next lesson · 12 of 14

Latency changes how a take feelsA few milliseconds of monitoring delay can make a good player sound stiff. Track at a small buffer or with direct monitoring, and raise the buffer again to mix.

Previous: What a correlation meter tells you

Portrait of Virzy Guns

Written by Virzy Guns

Songwriter and producer, founder of Virzy Guns Production. Now building HealingWave.