algorithmic vs convolution reverb
the two ways plugins create reverb: sampling a real space with impulse responses, or simulating one with delay networks. how each works, what each costs, and when to reach for which.
a photograph or a painting
every reverb plugin you have ever opened does one of two things. it either plays your signal through a recording of a real space, or it builds a synthetic space out of delays and feedback, one reflection at a time.
the first is convolution. the second is algorithmic. the distinction sounds academic until you realize they fail in opposite directions, and that most reverb frustration comes from asking one type to do the other one’s job.
a convolution reverb is a photograph of a room. perfectly accurate, completely frozen. an algorithmic reverb is a painting of a room. never exactly real, endlessly revisable.
convolution: the sampled space
a convolution reverb starts with an impulse response: a recording of how a real space answers a single, instantaneous sound. fire a starter pistol in a church, record what comes back, and you have captured every reflection that church produces. their exact arrival times, their levels, their tone, the way the high end dies faster than the low end.
convolution then applies that fingerprint to your audio. every sample of your signal triggers the entire recorded reflection pattern, scaled to that sample’s level, all of them summed. your snare is now, mathematically, being played back through that church.
the realism is the whole point. if you need the vienna konzerthaus, or the exact plate that lives in abbey road, convolution gets you that and nothing else does.
the frozen-ness is the cost. the IR is a finished recording. you can fade it shorter, EQ it, stretch it slightly with artifacts. what you cannot do is make the room behave differently: a longer decay, a wider stereo image, a tail that moves. the photograph does not let you walk around the building.
why convolution needs the FFT
naive convolution multiplies every input sample against every sample of the IR: a 5-second IR at 48 kHz is 240,000 multiplications per output sample, per channel. real-time convolution is only practical because of fast FFT-based convolution, which does the work in the frequency domain, and partitioned schemes that chop the IR into blocks to keep latency low. it is an elegant solution to a brute-force problem, and the cost still grows with IR length.
algorithmic: the simulated space
an algorithmic reverb owns no recordings. it builds the three stages of a real space (direct, early, late) from scratch, using networks of delay lines, feedback, and filters.
the lineage is worth knowing because every modern algorithmic reverb still carries it. schroeder showed in 1962 that combs and allpass filters could fake a tail.[^1] moorer added the missing pieces of realism in 1979: denser early reflections and air absorption.[^2] jot’s feedback delay network (FDN) work around 1991 gave the field its modern backbone: a handful of delay lines feeding back into each other through a mixing matrix, so every reflection re-excites every delay line, and echo density snowballs the way it does in a real room.[^3]
here is the thing about a simulated space: because every parameter is a number the designer chose, every parameter can move. decay time is a filter gain. room size is a set of delay lengths. the tail can be modulated, frozen, made to bloom or breathe. none of this is possible with a photograph.
the cost is the opposite of convolution’s. an algorithmic reverb will never be the konzerthaus. badly designed, it will not even be a plausible room: too few delay lines or carelessly chosen lengths and the tail rings at audible pitches, the metallic sound that stage 3 of this path dissects in detail.
key takeaway
convolution gives you a real space you cannot change. algorithmic gives you a changeable space that was never real. pick by which failure you can live with on this particular sound, not by which type is “better.”
when to reach for which
in practice the choice is less philosophical than the theory suggests. (cooking has the same split, for what it is worth: a stock cube is a recording of a kitchen’s work, frozen and consistent. a stock you simmer yourself is adjustable the whole way. nobody argues one is universally correct.)
reach for convolution when the space itself is the point. post-production dialogue that must sit in a specific real room. orchestral mockups that need a known hall. matching an overdub to the room the original was tracked in.
reach for algorithmic when the space serves the music. mix reverb that needs its decay tuned to the tempo. tails that should sound bigger than any real building. anything frozen, infinite, modulated, or evolving. sound design where the room is an instrument.
reach for algorithmic when CPU matters and tails are long. the cost of a delay network does not grow with decay time. a sixty-second tail costs what a two-second tail costs.
and a quiet third option: many engineers use both at once. a short convolution for the early reflections (the part where realism is cheap and matters most) feeding an algorithmic tail (the part where flexibility matters most). the photograph for the walls, the painting for the air.
note
a quick test that tells you which type you are listening to: turn the decay control, if there is one, through its whole range. a space that stays plausible at every setting is algorithmic. a control that mostly fades or truncates the same fixed tail is convolution wearing an algorithmic costume.
the algorithmic side is what i have been building. OPEN is an algorithmic reverb made for synths and drum machines rather than another plate emulation, and it opens for beta soon. if you want first access, tick “i want to beta test future plugins” on the signup at /#signup.
frequently asked questions
frequently asked questions
what is the difference between algorithmic and convolution reverb?
convolution reverb plays your signal through a recorded fingerprint of a real space (an impulse response), so you get that exact room, exactly as it was captured. algorithmic reverb builds the reflections from scratch using networks of delays and feedback, so the space is synthetic but every property of it is adjustable. convolution is a photograph; algorithmic is a painting.
what is an impulse response?
an impulse response (IR) is a recording of how a space reacts to a single, instantaneous burst of sound: a starter pistol, a popped balloon, or a swept sine wave that gets deconvolved afterwards. it contains every reflection the room produces, with its exact timing, level, and tone. convolving your signal with the IR applies all of those reflections to every sample of your audio.
is algorithmic or convolution reverb better?
neither is better; they fail in opposite directions. convolution is unbeatable at realism for a fixed, real space, but the space is frozen: you cannot make the captured room 20 percent longer or wider, only fade or filter what was recorded. algorithmic reverb never sounds exactly like a specific real hall, but the decay, size, tone, and modulation are genuinely adjustable, it can run lighter on CPU, and its tail can evolve in ways no static recording can.
why do algorithmic reverbs sometimes sound metallic?
a delay network with too few delay lines, poorly chosen delay lengths, or no modulation produces reflections that repeat in audible patterns. the repeats stack into resonances at related frequencies, and the tail takes on a pitch: the classic metallic ring. well-designed algorithmic reverbs avoid this with mutually prime delay lengths, dense diffusion, and slow modulation of the delays.
which reverb type uses more CPU?
it depends on the length. convolution cost grows with the IR: a 10-second cathedral IR is a lot of multiplication, even with FFT-based fast convolution, and long IRs add either latency or significant complexity to avoid it. an algorithmic tail costs the same per sample whether it rings for one second or sixty, which is why infinite and freeze tails are only practical algorithmically.
references
a note from the developer
i used to assume convolution was simply the superior technology and algorithmic reverbs were the budget option. it is such an easy story: real room, real physics, of course it wins. then i actually read the schroeder and jot papers, and the thing that struck me was that the algorithmic designers were never trying to imitate convolution. they were chasing a different goal from the start: a space you can steer. sixty years of research into making the painting respond like a room, precisely because a photograph cannot.
these days when i hear a tail i love on a record, my first question is not “what hall is that”. it is “what is that tail doing that no hall would do”. the answer is usually the interesting part.
if you disagree with any of this, or you have a convolution-vs-algorithmic war story, my inbox is open. jonas@kernaudio.io. i read every email.