Reverb Deep Dive

Space As A Filter

A room is a linear time-invariant filter with a several-hundred-thousand-tap impulse response. Everything on this page is an attempt to build that filter out of something cheaper than a room: a coiled spring, a sheet of steel, four kilowords of RAM, an orthogonal matrix, a measured recording, or a set of gradients. The engineering history of reverberation is the history of those substitutions, and each one leaves a signature you can still identify by ear.

Three families exist, and they have never stopped competing:

                     the three ways to make a space

  MECHANICAL               ALGORITHMIC                CONVOLUTIONAL
  ─────────────            ──────────────             ──────────────
  build a physical         build a recursive          measure a real
  resonator with high      network of delays          impulse response
  modal density            that grows echoes          and convolve with it
  │                        │                          │
  chamber (1947)           comb + allpass (1961)      Sony DRE-S777 (1999)
  spring  (1939)           FDN (1971 / 1982)          Altiverb (2001)
  plate   (1957)           allpass loop (1992)        partitioned FFT
  tape    (1950s)          scattering net (2015)      hybrid ER + FDN tail
  │                        │                          │
  cost: space, weight      cost: coloration           cost: no parameters
  win:  free physics       win:  tweakable            win:  it is the room

  and the fourth, which is really the first: SIMULATION ─ solve the wave
  equation or trace the rays, then render. Games and architecture live here.

The one number that organises everything

Reverberation time, T60, is the time for the sound pressure level to fall 60 dB after the source stops. It is the axis on which every device on this page is specified, tuned, and marketed. W. C. Sabine established it empirically at Harvard between 1895 and 1900 while trying to fix the Fogg lecture hall, using organ pipes, a stopwatch and his own ears, and the formula he arrived at is still the first thing anyone computes.

Sabine's formula assumes a perfectly diffuse field and low absorption. When absorption is high it overestimates, and the Norris-Eyring correction replaces the linear absorption sum with a logarithmic one. For a live room the difference is a few percent; for a treated control room it is large.

Worked example, verified numerically: a tracking room 9 × 7 × 4 m has V = 252 m³ and S = 254 m². At a mean absorption of 0.18, Sabine gives T60 = 0.89 s and Eyring gives 0.80 s, a 9 percent disagreement. That gap is why measurement, not calculation, settles arguments about rooms.

The Wave Equation And Its Consequences

Everything downstream is an approximation to one second-order PDE. Sound in a lossless fluid at rest obeys it, and reverberation is what its solutions look like inside a closed boundary.

Impose rigid walls (normal particle velocity zero, so ∂p/∂n = 0 at every surface) on a rectangular box and the solutions quantise. You get standing waves at discrete frequencies: the room modes.

For the 9 × 7 × 4 m room, the lowest axial modes sit at 19.1, 24.5 and 42.9 Hz. Those are the frequencies that make small rooms untrustworthy at the bottom end, and no amount of algorithmic reverb fixes them because they are in the monitoring path, not the signal.

Modal density and the Schroeder frequency

Count how many modes exist below frequency f and differentiate. Modal density grows as f², which means that at some point the modes overlap so thoroughly that talking about individual resonances stops being useful.

The crossover is the Schroeder frequency, the point at which roughly three modes fall inside one modal bandwidth. Below it the room is a set of resonators and you should think in modes. Above it the room is a stochastic diffuser and you should think in statistics. This single boundary is why reverb algorithms and room-correction algorithms are different disciplines.

Mean free path and echo rate

Two geometric statistics do most of the work in algorithmic design. The mean free path is the average distance a ray travels between reflections; its reciprocal in time gives the average reflection rate.

The frequency-domain view

The transfer function of a room is a dense forest of poles. Each mode contributes a complex-conjugate pole pair with half-power bandwidth set by its damping. Because the poles are packed far tighter than their bandwidths above f_s, the magnitude response looks like filtered noise: a Rayleigh-distributed magnitude with roughly 5.6 dB standard deviation, and a mean spacing between response maxima of about 4/T60 Hz. Any algorithm whose magnitude response has visible periodic structure will read as "metallic", and that is a measurable defect, not a matter of taste.

Anatomy Of An Impulse Response

Every reverb, mechanical or digital, is judged against the shape of a measured room impulse response. Learn to read this picture and most design decisions become obvious.

The three regions, and what each one is for

1. Direct sound, t = 0The unreflected wavefront. Carries source identity, timbre and localisation. Not part of reverb, but its level relative to everything after it (the direct-to-reverberant ratio) is the primary distance cue. Turning up a reverb send is, perceptually, walking away from the source.
2. Early reflections, roughly 1 to 80 msDiscrete, sparse, specular. Geometric acoustics is valid here, so each reflection has a computable path, arrival time, direction and spectral tilt from the surfaces it bounced off. These reflections do not sound like reverb; they sound like place. They set apparent room size, source distance, and, through interaural differences, apparent source width. The initial time delay gap (ITDG) between direct and first reflection is the strongest single size cue: 5 ms says small room, 40 ms says cathedral. In a mixer this is the pre-delay control, and it is the most underused parameter on any reverb.
3. Late field, roughly 80 ms onwardEcho density has passed a few thousand per second, individual reflections are unresolvable, and the signal is statistically indistinguishable from Gaussian noise multiplied by a decaying exponential envelope. Frequency-dependent: air absorption and surface absorption both rise with frequency, so the high band decays faster and the tail darkens as it fades. This region is where all the DSP cleverness goes, because it is 99 percent of the samples and 100 percent of the CPU.

Mixing time

The boundary between region 2 and region 3 is the mixing time: the point at which the field becomes diffuse. A useful engineering rule for a roughly cubic room is t_mix ≈ √V milliseconds with V in m³, giving about 16 ms for our tracking room and 120 ms for a large hall. Below t_mix, model geometry. Above it, model statistics. Every serious modern reverb, from TC Electronic's VSS algorithms to a game engine's acoustic renderer, splits its architecture at exactly this line.

Reading decay: Schroeder backward integration

You cannot read T60 off a raw impulse response, because the tail is noise and the instantaneous amplitude fluctuates wildly. Schroeder's 1965 trick is to integrate the squared IR backwards in time, which converts a single noisy measurement into the ensemble-average decay curve you would otherwise need many measurements to estimate.

The same paper family gives you the two standard measurement excitations. Maximum length sequences (MLS) are deterministic pseudorandom binary sequences whose circular autocorrelation is a delta; you cross-correlate the response with the sequence and get the IR with a large SNR gain. The modern default is Farina's 2000 exponential sine sweep: sweep logarithmically from 20 Hz to 20 kHz, deconvolve with a time-reversed amplitude-corrected copy of the sweep, and the harmonic distortion products land at negative time where you can simply crop them away. That property is why sweeps beat MLS on real loudspeakers, which are never quite linear.

Rooms As Machines

The first artificial reverb was a real room used deliberately. Dedicated reverberation chambers, loudspeaker at one end and microphone at the other, were radio-network practice by the 1920s (the oft-repeated claim that RCA patented the chamber in 1926 has no locatable patent behind it), but the technique became a production tool in 1947 in a bathroom in Chicago.

The echo chamber, 1947

Bill Putnam Sr, working at Universal Recording in the Chicago Civic Opera Building, put a loudspeaker and a microphone in the tiled women's washroom and mixed the return into The Harmonicats' recording of Peg o' My Heart. Contemporary accounts have engineer Bernie Clapper on the build and somebody posted at the door to stop anyone flushing mid-take. The record went to number one and sold over a million copies, and the send-return architecture of every mixing console since descends from it: dry signal on the fader, wet signal on a separate return, ratio under the engineer's control.

Purpose-built chambers followed immediately, and their design rules are worth stating because they are the acoustic constraints the digital algorithms later had to satisfy:

• No parallel surfaces. Parallel walls give strong axial modes and flutter echo. Trapezoidal plans and irregular ceilings scatter instead.
• Hard, sealed, non-absorbing finishes. Concrete sealed with lacquer, glazed tile, plaster. Long T60 requires α to be tiny, and Sabine's formula tells you exactly how tiny.
• Small volume, long decay. A short mean free path builds echo density fast (high reflection rate) while low absorption keeps the tail long. That combination is the chamber sound: immediate onset, no perceptible ITDG, dense from the first millisecond, and it is very hard to fake with a sparse early-reflection model.
• Mechanically isolated. The chamber is a microphone. Traffic, HVAC and footsteps all print.

The famous installations: Capitol Studios in Hollywood, concrete trapezoidal chambers roughly 30 feet under the Tower's parking lot, designed by acoustician Michael Rettinger (building on Les Paul's earlier reverb experiments), with decay times to about five seconds, four chambers in the original 1956 build and four more added in 1969; Abbey Road, one chamber per studio, later wired into the STEED send chain that put tape delay in front of the chamber; Gold Star, whose two chambers are structural to Phil Spector's Wall of Sound; and Motown's Hitsville attic, an untreated space whose accidental character defined a label.

Tape and oil-can delay: the adjacent lineage

Tape loops with multiple playback heads (Echoplex, Roland RE-201 Space Echo, Binson Echorec with its magnetic drum) are not reverb, but they occupy the same send slot and they contributed one crucial idea: feedback around a delay. Route the playback head back into the record head at gain g and you have built, in iron and oxide, the feedback comb filter that Schroeder had described mathematically. The Binson's four heads at fixed spacings on a rotating drum is literally a four-tap delay line with recirculation, and its wow, flutter and head-bump filtering are the analogue ancestors of the modulated delay lines that later became the Lexicon sound. Oil-can delays (Tel-Ray, using a rotating dielectric fluid capacitor) got there by an even stranger route.

Springs · A 1D Dispersive Waveguide

The idea is to trade three dimensions for one, and buy back the missing modal density with a medium in which waves travel slowly and, crucially, at a speed that depends on frequency.

Provenance

R. L. Wegel at Bell Labs patented torsional-wave delay in a coiled spring (US 1,852,795) as telephone research. Laurens Hammond adapted it for musical reverberation, filing in 1939 (granted as US 2,230,836, 4 February 1941) with the explicit goal of "introducing a reverberation effect of selected degree in the music irrespective of the acoustic properties of the place where the instrument is being played". The market was churches whose sanctuaries had been acoustically deadened so the sermon would be intelligible, which left the organ sounding dry. Hammond's Type 4 tank and its successors were productised at scale by Hammond's spin-off and later Accutronics; the canonical engineering description is Meinema, Johnson and Laube, "A New Reverberation Device for High Fidelity Systems", JAES 9(4), 1961, which is also where the lumped-element model comes from.

The electronics

The physics of the drip

A helical spring supports several wave types that couple into each other: transverse (the coil swinging sideways), longitudinal (coils compressing along the axis) and torsional (each coil twisting about the wire axis). Reverb tanks are driven and read torsionally. The essential fact is that the medium is strongly dispersive: phase velocity depends on frequency, so an impulse does not stay an impulse. It smears into a chirp.

Model the coil as a one-dimensional lumped lattice, one torsional compliance and one inertance per turn. For a spring of M coils you get M modes, and the dispersion relation has a hard cutoff:

The other property engineers exploit is that a spring tank is only approximately linear. Hit it hard and the transverse mode clips against the enclosure while coils collide in the longitudinal mode, which cross-couples into torsion as a sputtery burst. This is the surf-guitar reverb crash, and it is a nonlinearity, so no LTI convolution of a spring IR will ever reproduce it. Neural and Volterra-series models exist precisely to capture that.

Plates · A 2D Bending Sheet

Go from one dimension to two and modal density stops being a problem. A thin plate's modal density is constant with frequency (mode count grows linearly, where a room's grows as f³), and the constant is enormous because the plate is large and thin, so a 2 × 1 m steel sheet has thousands of modes in the audio band and produces a dense, flutter-free tail that no spring can match.

Provenance, corrected

The plate reverberator was developed by Dr Walter Kuhl at the Institut für Rundfunktechnik (German broadcasting's research institute) in the mid-1950s. EMT, Elektromesstechnik Wilhelm Franz KG of Lahr in the Black Forest, brought it to market in 1957 as the EMT 140 and held the patents. Wilhelm Franz founded and ran EMT; he is not the plate's inventor, a conflation that appears in a lot of secondary writing. The stereo version, the 140 ST with a second pickup, arrived in 1961.

Mechanics

• The medium. Transverse bending waves in a thin plate, not longitudinal waves in air. Bending-wave propagation in a plate is also dispersive, with phase velocity proportional to √ω, but the modal density is so high that the dispersion reads as a bright shimmer rather than as discrete chirps.
• Transduction in. A moving-coil driver bonded to the centre of the plate, working normal to the surface. Audio current in the coil launches bending waves radially.
• Transduction out. Two piezoelectric contact pickups at deliberately asymmetric distances from the exciter. The asymmetry decorrelates the two channels, giving pseudo-stereo from a single mono excitation, and no two 140s were identical because the geometry was hand-set.
• Decay control. A parallel damping panel of framed acoustic tiles on motorised arms. It never touches the plate; it changes the air-coupled radiation load, and moving it scales T60 continuously from roughly five or six seconds down to about half a second. This is the first artificial reverb with a decay knob, and that is why it took over from chambers.
• Frequency response. A rise at the bottom and a decline at the top, reaching about 1.5 s at 10 kHz, which mimics air absorption without anyone having designed it in. Not a coincidence, but not deliberate either: the plate radiates high frequencies more efficiently, so they lose energy faster.

The 240 gold foil, and the digital 144

EMT's follow-up, the EMT 240 of 1971, replaced the two-metre steel plate with an electrolytically produced gold foil of 270 × 290 mm, dropping the enclosure to about one fifth the volume and, critically, removing the need to re-tension after transport. It reads brighter and shimmerier than the 140 and covers about 0.5 to 5 s of decay, in true stereo. There was also an EMT 144 in 1972, an early digital attempt that did not sell and barely survives, which makes EMT one of the few companies that shipped chamber-replacement, plate, foil and digital reverberators.

The other plate-family entries: AKG's BX 20 and BX 15, which are torsional spring systems in a floor-standing cabinet rather than plates, often grouped with plates because of their size and studio role; and the various Grampian and Fisher spring units that put reverb into consumer hi-fi.

Schroeder's Two Primitives

Manfred Schroeder at Bell Labs published the founding papers: "Colorless" Artificial Reverberation (with B. F. Logan, JAES, 1961) and Natural Sounding Artificial Reverberation (JAES, July 1962). Between them they establish that reverberation can be synthesised from two discrete-time building blocks, and every algorithmic reverb since is assembled from these or their descendants. Note that these were offline computations on a mainframe: seconds of audio took hours. Real-time was fifteen years away.

Primitive 1 · the feedback comb filter

A delay of m samples with its output fed back at gain g. It emits an infinite, strictly periodic pulse train.

Primitive 2 · the allpass filter

Wrap the same delay in a feedforward path with gain -g and the magnitude response becomes exactly flat while the phase, and therefore the group delay, still disperses. Schroeder's insight in 1961 was that this lets you increase echo density without adding coloration, which is the central trick of the whole field.

practice# Dattorro / Gardner two-multiply form, one buffer, equivalent up to a sign on g:
d = buf.read(m)
t = x[n] - g*d      # write t into the buffer
y[n] = d + g*t

The 1962 topology

Four parallel comb filters with mutually prime delay lengths, summed, then two allpasses in series. The parallel combs give decay and a first pass at density; prime lengths ensure their peak patterns do not align, so the combined ripple partially averages out. The series allpasses multiply the echo count without adding further ripple.

Why it is not enough

Three failures, all audible:

• Constant echo density. Four combs give four fixed pulse rates that never accelerate. Real rooms grow density as t². The result is a rattly, granular tail that reads as small and metallic no matter how long you set the decay.
• Residual periodic ripple. Prime delays decorrelate the peak patterns but do not eliminate them, and with only four combs the summed magnitude response is still visibly periodic. The classic symptom is a fixed set of pitches ringing out of the tail.
• Frequency-independent decay. Every mode decays at the same rate, so the tail stays as bright at three seconds as at ten milliseconds. Real air and real surfaces absorb high frequencies preferentially. This is the specific defect Moorer fixed.

Two Lineages Diverge, 1971 to 1979

Within a decade of Schroeder, the field split into two philosophies that are still recognisable in products today. One extends the cascade: more allpasses, more filters, tuned by ear, aimed at a beautiful result. The other replaces the cascade with a matrix: a network of delays cross-coupled by a unitary transform, aimed at a correct result. The commercial reverbs you know inherit from one or the other, and often from both.

Branch A · Moorer, 1979: fix the decay spectrum

James A. Moorer's About This Reverberation Business (Computer Music Journal, 1979) is the practical engineering follow-up to Schroeder. Three contributions:

1. The lowpass comb filterPut a one-pole lowpass inside the feedback loop. Now each circulation attenuates high frequencies more than low, so T60 becomes frequency dependent, which is what air absorption and surface absorption actually do. This single change is the difference between "digital reverb" and "digital reverb that sounds like a room".
2. The 2-multiply allpassA more efficient allpass structure than Schroeder's direct form, which became the standard building block and is what almost every implementation uses today.
3. An explicit early-reflection FIRA sparse tapped delay line in front of the recursive network, with tap times and gains derived from room geometry, feeding the diffuse network. This is the first appearance of the split architecture (geometric early field, statistical late field) that every serious reverb since has used.

Branch B · Gerzon, 1971: replace the cascade with a matrix

Michael Gerzon, writing in Studio Sound (December 1971 and January 1972), described what we now call the feedback delay network, ten years before it entered the American literature. His observation was blunt: a single feedback comb sounds poor, but several cross-coupled combs can sound excellent. So put N parallel delay lines in a loop and close the loop through an orthogonal matrix.

Because the matrix is energy-preserving, the network is stable and lossless by construction, and because every delay feeds every other delay, the echo count multiplies on each circulation: density grows without bound, which is exactly the t² behaviour real rooms have and the Schroeder topology lacks. Gerzon also showed how to build higher-order unitary matrices out of 2 × 2 rotations, how to use the rotation angles to control the stereo spread of the reverberation, how to insert filters inside the network to get frequency-dependent decay, and how to replace the allpass multipliers with filters while preserving the allpass property. In 1976 he formalised the theory in Unitary (energy preserving) multi-channel networks with feedback (Electronics Letters 12(11)).

A third strand runs alongside: Julius O. Smith's A New Approach to Digital Reverberation using Closed Waveguide Networks (ICMC 1985) builds reverberators from bidirectional waveguides joined at scattering junctions. It matters because it connects reverb to physical modelling proper, and because the waveguide junction is a scattering matrix, which is where later FDN designs get many of their feedback matrices from.

Feedback Delay Networks

The FDN is the workhorse of modern algorithmic reverb. N delay lines in parallel, their outputs mixed by an N × N feedback matrix and returned to their inputs. It generalises the comb filter (which is the N = 1 case with A = [g]) and, with the right matrix and the right per-line filters, it subsumes most other topologies including Schroeder's cascade, Moorer's extension and Dattorro's allpass feedback network.

Losslessness: what makes it stable

The design discipline is to build a lossless prototype first (a network that rings forever without growing), then add attenuation to set decay. If A is unitary (orthogonal in the real case, so AᵀA = I), the total energy circulating in the delay lines is exactly conserved on every pass through the matrix and the network is lossless. All eigenvalues of A then lie on the unit circle.

Setting decay: Jot's absorbent filters

Naively, multiply the whole matrix by a scalar g and everything decays. The problem is that different eigenmodes of the network then decay at different rates, and the slowest-decaying ones ring on as isolated pitches: metallic coloration. Jot and Chaigne's 1991 fix is to give each delay line its own attenuation, scaled so that every line loses the same amount of energy per unit time rather than per circulation.

Verified: an N = 4 FDN at 48 kHz with a normalised Hadamard matrix, delays m = [997, 1153, 1327, 1559] and per-line gains from the formula above measures T60 = 2.002 s against a 2.000 s target using T30 backward integration; per octave band, 2.005 s at both 200 Hz and 4 kHz. The formula is exact, not approximate.

Tonal correction

One more filter completes the classic design. Shortening the decay in a band also removes energy from that band in the impulse response, so a reverb with a short high-frequency T60 sounds dull in a way that is not physical. Jot's tonal correction filter E(z) sits at the network output with a gain that rises where decay is short, equalising the total energy per band irrespective of the decay time. It is the difference between a decay-time control that changes the decay and one that also changes the timbre.

Density, delay choice, and modulation

• N sets the density floor. An N = 4 FDN with delays near 1000 samples emits only 8 nonzero taps in the first 50 ms at 48 kHz (verified). That is not diffuse. The standard fixes are to raise N to 8 or 16, to precede the FDN with a diffusion stage of series allpasses, or to embed allpasses inside the delay lines so each "delay" is itself a diffuser.
• Delay lengths. Mutually prime, spread over roughly a 1:1.5 to 1:3 ratio, and long enough that the shortest is above the room's expected first reflection time. A widely used set at 48 kHz is [997, 1153, 1327, 1559, 1801, 2099], all prime.
• Modulation. Slowly vary the delay lengths (or rotate the feedback matrix) and the network becomes time-varying. Standing modes cannot establish, residual ringing is smeared, and the tail acquires a chorus. This is the Lexicon sound, and it is also why Lexicon reverbs are not phase-coherent and why classical engineers historically preferred TC Electronic.
• Sparsity. A shifted-identity feedback matrix makes the FDN degenerate into a single long delay loop; the identity matrix makes it N independent combs. Everything interesting lives between those two extremes.

Allpass Loops And Nested Diffusion

The third major topology, and the one that actually shipped in the most famous hardware, is the allpass loop: a chain of allpass filters embedded inside a single larger delay loop, with the loop gain setting the decay. It circulated privately in the electronic instrument industry through the 1980s and reached the open literature only in the 1990s.

Nested allpass

The construction that makes it work is the nested allpass: replace the delay inside an allpass with another complete allpass. The result is still exactly allpass, but its impulse response is far denser than either component. Gerzon described the principle in 1972; Bill Gardner's A Realtime Multichannel Room Simulator (ASA, November 1992) is the first public treatment with concrete designs, and it is also where the single-shared-buffer implementation technique comes from.

The allpass loop

Jon Dattorro's Effect Design, Part 1: Reverberator and Other Filters (JAES 45(9), September 1997) published a complete working reverberator of this type, "in the style of Griesinger", with every delay length, every coefficient and every output tap listed. The effect on the field is hard to overstate: that one paper is the origin of an enormous fraction of the plate reverbs in commercial plugins, in synth effects sections and in game engines, often reproduced tap for tap. The paper also carries commentary from Barry Blesser explaining why delay modulation went into commercial reverbs in the first place, which makes it a primary historical source as well as a design document.

Modulation, and the Lexicon signature

David Griesinger's contribution at Lexicon was to make the network time-varying. Delay tap positions inside the loop are continuously modulated, in the 224 by low-frequency oscillators and in the 480L's Random Hall by pseudorandom tap displacement. Two effects:

• Pole smearing. A time-varying delay length means the network's poles are never stationary long enough for a mode to establish. Ringing that would be a fixed pitch becomes a wandering band, which the ear integrates as smoothness rather than hearing as a tone.
• Higher usable feedback. Christopher Moore's Ursa Major patent makes the same point from the other direction: modulating the taps let him run roughly three times the feedback gain before the structure became unstable or obviously periodic. Time variation buys stability margin.

The cost is pitch modulation. Every reflection inside the loop is being slightly resampled, so sustained tonal material acquires a chorus and phase coherence between channels is not preserved. For pop and film that reads as lush and expensive. For classical recording and mastering it is a defect, which is exactly the market TC Electronic went after with strictly time-invariant designs.

The Iron Age · 1976 to 2007

Schroeder's algorithms ran offline on Bell Labs mainframes: seconds of audio, hours of compute. Real-time reverberation needed a computer that was fast at multiply-accumulate, cheap enough to sell, and equipped with enough RAM to hold a few hundred milliseconds of audio. In 1976 all three of those were barely true at once, and the machines built at that boundary have characters that people still pay for.

EMT 250, 1976 · the first commercial digital reverb

EMT, already the incumbent in mechanical reverb with the 140 and 240, went digital by contracting the work out. The algorithms were designed by Dr Barry Blesser of MIT together with EMT's technical director Karl-Otto Bäder (US patent 4,181,820); the digital hardware came from Ralph Zaorski at the Massachusetts contract engineering firm Dynatron / Dynatronics; EMT built the converters, I/O, power supply and the enclosure. Bäder introduced it at the Zürich AES convention in 1976. Peter Bermes is also credited on the project, and the industrial design (the four upright levers, the red external power supply) was contributed by a freelance designer, which is why the machine looks like a prop.

          EMT 250 "the Robot" / "Spaceheater" · 1976 · $20,000

  ┌────────────────────────────────────────────────────┐
  │              EMT 250 front panel                   │
  │                                                    │
  │    ▐▌        ▐▌        ▐▌        ▐▌                │  large upright
  │    ▐▌        ▐▌        ▐▌        ▐▌                │  mechanical levers,
  │    ▐▌        ▐▌        ▐▌        ▐▌                │  linear pots, an
  │  DECAY    HF-DAMP   LF-DAMP   PREDELAY             │  explicit visual
  │                                                    │  quote of the 140's
  │  programs: REVERB · DELAY · PHASE · CHORUS · SPACE │  damping handwheel
  └────────────────────────────────────────────────────┘

  inside   ~400 discrete TTL ICs, no DSP chip existed yet
           16 kwords of RAM (1 kbit chips), ≈ 400 W, three fans,
           heat sinks on the outside of the cabinet
           12-bit converters, quasi-15-bit, ≈ 24 kHz sample rate
           1 input, 4 outputs · about 250 units built

  algorithm  one large, heavily cross-coupled allpass network, designed
           to consume every available word of RAM without producing
           discrete echo nodes. Blesser's team built a programmable
           simulator to audition candidate algorithms, roughly two
           years of R&D. Delay-line modulation was in from the start.

Lexicon 224, 1978 · the one that defined the sound

Lexicon began in 1969 as American Data Sciences, founded by MIT professor Dr Francis F. Lee and engineer Chuck Bagnaschi to make digital delay lines for medical use (ECG and heart-sound analysis); renamed Lexicon in 1971, it pivoted to pro audio, with Blesser involved in its early days. The 224 came from outside: David Griesinger, a nuclear physicist, musician and classical recording engineer, had been experimenting with digital reverberation, saw the EMT 250, combined his algorithm work with microcomputer control, and pitched a prototype. Lexicon bought the design and brought him in to finish it. Announced at the 1978 AES convention; contemporary trade press lists $7,500 for the two-program version and $7,900 with four programs (later sources say $8,800).

                  Lexicon 224 · architecture as built

   analog in
      │
      ▼
  ┌────────────────────────┐   gain-ranging ADC: a 12-bit linear
  │ 12-bit + analog gain   │   converter (DAC80 family) preceded by
  │ ranging, 6 dB steps    │   an automatic gain-ranging stage that
  └───────────┬────────────┘   shifts level in discrete steps and
              │                stores the exponent. Buys roughly
              ▼                24 dB of extra headroom → pseudo-
  ┌────────────────────────┐   floating-point, ≈ 90 dB range.
  │ 16-bit fixed-point     │   The step transitions are part of
  │ engine, Intel 8080     │   the grit people pay for.
  │ control CPU + a great  │
  │ deal of 74S/74LS logic │   Dynamic RAM for the delay buffers,
  │ + dynamic RAM buffers  │   the scarcest and most expensive
  └───────────┬────────────┘   resource in the whole design.
              │
              ▼                4U mainframe + separate desktop
  ┌────────────────────────┐   remote. Griesinger insisted on the
  │ 12-bit DAC + ranging   │   split so the engineer had parameter
  └───────────┬────────────┘   access at the console; the 224X and
              ▼                224XL later got the LARC remote.
   analog out

  the algorithm  allpass loops with continuously modulated delay taps.
  Sub-audio LFOs shift the internal tap positions, pitch-shifting the
  recirculating reflections slightly, which prevents standing modes and
  produces the smooth chorused tail known as the Lexicon sound. Split
  bass/mid decay with a crossover, plus separate treble decay, gave a
  tunable decay spectrum years before anyone published the theory.

Ursa Major Space Station SST-282, 1978 · the cheap outsider

Christopher Moore was project engineer at Lexicon; he resigned in 1977, founded Ursa Major, and set out to build a good digital reverb for $2,000 rather than the price of a 224 (it shipped at $1,995, launched at the Los Angeles AES convention in May 1978; roughly 1,900 were made). The architecture is not a Schroeder cascade. It is a recirculating multitap delay: a single 255 ms buffer with eight audition taps whose positions float, moving randomly under program control. Moore described it in the manual as a special multi-head tape recorder with a 255 ms loop.

The floating taps are the invention, and they are patented: randomising tap positions is what lets the structure run at high feedback without either self-oscillating or turning into an audible periodic echo. Moore later catalogued its flaws honestly (spectral smearing from wandering taps, modulation noise from unsmoothed 62 µs tap jumps, and an inability to sound truly distant because the audition taps picked up dry source as well as diffuse tail). Those flaws are the sound. Ursa Major was bought by AKG in 1986; Moore's later designs (the 8×32, the AKG ADR-68K, algorithms for Kurzweil) used stable rather than time-variant structures.

AMS RMX16, 1982 · microprocessor control and the gated preset

Advanced Music Systems, founded in 1976 in Burnley by ex-aerospace engineer Mark Crabtree with Stuart Nevison, built the DMX 15-80 digital delay in 1978, then the DMX 15R reverb expander in September 1981, then repackaged that combination as the self-contained RMX16 in March 1982: a microprocessor-controlled digital reverberator running a 16-bit engine behind a 12-bit gain-ranged converter, physically compact next to a 250 or a 224, with a display and stored programs.

Its historical importance is a preset. In 1979 at Townhouse Studio 2, engineer Hugh Padgham was setting up drums for Peter Gabriel's third album when he engaged the reverse talkback on the studio's new SSL 4000 B console while Phil Collins was playing. The talkback mic was heavily compressed so that quiet speech from the live room would be audible; run through the console's per-channel gate as well, it produced an enormous room sound that snapped to silence. It went on "Intruder", then on "In the Air Tonight". AMS, working closely with Townhouse, turned the effect into the RMX16's Non Lin 2 preset: a deliberately unphysical reverb whose envelope does not decay, and in places rises, before terminating abruptly. Gated reverb as a preset rather than a signal-chain trick is the single most identifiable production signature of the 1980s.

Quantec QRS, 1982 · the outlier that modelled propagation

Wolfgang Buchleitner (born Wolfgang Schwarz), a self-taught German designer with no engineering degree, founded Quantec in Munich in 1982 and shipped the Room Simulator. Presented at the Montreux AES show, it took a different route from every competitor: rather than assembling echo and feedback networks, the QRS modelled sound propagation and resonance in an enclosure. Eight parameters, decay times reported from about 0.1 s to 100 s (and several times that in the low bass), 16-bit conversion with oversampling and a wider internal word length, plus the first commercial Freeze function, which captured a signal in a non-parallel-walled virtual room and sustained it indefinitely.

It became the tool of choice for post-production ADR matching (its accuracy at reproducing a specific room was the selling point), and a texture generator for Peter Gabriel, Enya, Pink Floyd and Depeche Mode. Units were still being built in 1995; the Yardstick series continued the algorithm. Buchleitner died in 2016, and in November 2024 Apple acquired the Quantec technology and shipped faithful QRS and Yardstick recreations inside Logic Pro, working from Buchleitner's original schematics, algorithms and code. That is one of the very few cases of a proprietary reverb algorithm being preserved rather than reverse-engineered.

Lexicon 480L, 1986 · maturity

The 480L moved to 18-bit linear conversion at 44.1/48 kHz with digital I/O compatible with the Sony PCM-1610/1630, quoting 98 dB dynamic range in the wet path, and it ran on Lexicon's own microcoded arithmetic unit rather than an off-the-shelf DSP. Reverse-engineering work on the family describes a common ARU architecture across the Lexicon 200, 224XL, PCM60, PCM70 and 480L, with the 480 adding an extra T-state per instruction to fit one more multiply. The later single-chip Lexichip is essentially that ARU on one die, which is what made the cheap LXP series and later the PCM80/90 and the 300 possible; Lexichip II and III followed.

Algorithmically the headline was Random Hall: instead of periodic LFO modulation, the late-field tap positions are displaced by a random process. Random modulation has no cyclic artefacts at all, which is what let the 480L hold up on sustained piano and strings where the 224's chorus would have been obvious.

TC Electronic · phase coherence as a product strategy

TC Electronic was founded in 1976 in Risskov near Aarhus by brothers Kim and John Rishøj, starting with analogue guitar pedals (the SCF chorus) and moving up into high-end digital. The M5000 in 1994 and the System 6000 in 1999 pursued a philosophy explicitly opposed to Lexicon's: source purity, phase linearity, and geometric accuracy.

Their VSS (Virtual Space Simulation) family splits the problem cleanly:

Democratisation, 1985 to 1990

What actually changed music was not the flagships but the arrival of cheap single-chip DSP. The Yamaha SPX90 (1985) put usable reverb, delay and pitch shift in a 1U box at a few hundred dollars; the Alesis Midiverb and Microverb pushed the price under $400. Yamaha's REV-1 (1983) and REV-7, Roland's SRV-2000 and R-880, Eventide's SP2016 with its user-loadable algorithm ROMs, Sony's DRE-2000, EMT's own 245 and 251: within five years every project studio had a reverb, and the sound of records changed accordingly. The 1980s do not sound reverberant because engineers suddenly liked reverb; they sound reverberant because reverb became something you could buy for the price of a microphone.

Bricasti M7, 2007 · the last flagship box

Brian Zolner (Lexicon's long-time VP of international sales) and Casey Dowdell (a Lexicon DSP engineer) left to found Bricasti Design in 2004, the name a blend of their first names. The M7 is an algorithmic reverb with modern converters and a very large delay-memory budget, and it arrived precisely when conventional wisdom said hardware reverb was finished. It succeeded, and it remains the reference against which convolution and plugin reverbs are compared, which tells you something durable: with enough compute, a well-tuned algorithmic reverb beats a measured one, because it can be adjusted.

Convolution · Just Use The Room

If a room is an LTI system, its impulse response is a complete description. Record the IR, convolve, done. The theory has been obvious since the 1950s; the obstacle was arithmetic. A three-second stereo IR at 48 kHz is 144,000 taps per channel, so direct convolution costs 144,000 multiply-accumulates per sample per channel, roughly 14 GMAC/s for stereo in and stereo out. That is why convolution reverb arrived in 1999 and not 1979.

Measuring an impulse response

1. Impulse excitationPistol shot, balloon burst, clapper board, spark gap. Simple, but the SNR is whatever the impulse's energy gives you in one shot, the spectrum is uncontrolled, and loud transients are nonlinear. Still the only option for spaces you cannot bring a PA into.
2. MLS, maximum length sequencesA deterministic pseudorandom binary sequence of length 2ᴺ-1 whose circular autocorrelation is a delta. Play it, cross-correlate the recording with it (a fast Hadamard transform does this in N log N), and you get the IR with an SNR gain of roughly 10 log₁₀(L) dB for sequence length L. Because energy is spread over the whole sequence, peak level stays low. The fatal weakness: any nonlinearity in the loudspeaker scatters distortion products uniformly across the whole IR as a noise floor you cannot separate.
3. Exponential sine sweep (ESS)Farina, AES 108th Convention, 2000. Sweep a sine logarithmically from 20 Hz to 20 kHz over 10 to 30 s, then deconvolve with a time-reversed, amplitude-corrected copy. Because frequency advances exponentially, the n-th harmonic distortion product arrives at a fixed time before the linear response, so all distortion lands at negative time and you crop it off. You also get the harmonic distortion measurement for free from the same recording. This is the modern default, and the reason every commercial IR library sounds clean.
4. Post-processingCrop, denoise the tail (or replace the noise-buried portion with a synthesised extension), normalise, trim pre-delay, and decide on the direct sound: most libraries keep it, some strip it so the user controls dry/wet independently. Multichannel captures use spaced pairs, Ambisonic arrays or dummy heads.

Doing the arithmetic in real time

Frequency-domain convolution turns a 144,000-tap product into a pointwise multiply, but a single FFT of the whole IR would need 144,000 samples of input before it could produce anything: 3 seconds of latency. The solution is partitioned convolution: chop the IR into blocks, FFT each block once at load time, and for each incoming input block multiply and accumulate against all of them with the appropriate delays (overlap-save).

The hardware and the plugins

Sony's DRE-S777 (1999) was the first commercial real-time convolution reverberator: a rack unit with a CD-ROM drive for IR libraries and a memory card for patches, about £5,900 base and over £10,000 with the IR disc libraries. Sony's engineers travelled to real venues (the Concertgebouw among them) to capture them. Yamaha followed with the SREV1. TC Electronic had already been using limited real-time convolution for the early-reflection stage of VSS algorithms, which is worth noting as the first hybrid.

Then a chip made it free. When Apple shipped the PowerPC G4 in 1999/2000, its AltiVec SIMD vector unit made FFTs on a desktop cheap enough to convolve in real time. The Dutch company Audio Ease built Altiverb on it in 2001, and named it after the vector engine. Convolution reverb went from a five-figure rack to a plugin in two years, and the IR library became the product rather than the algorithm.

Hybrid: the current best practice

Because convolution is accurate but rigid and algorithmic reverb is flexible but approximate, high-end designs use both. Convolve a short measured early portion (the first 50 to 150 ms, which carries the room's identity and is cheap because it is short), then hand off to a parameterised FDN whose T60 curve is fitted to the measured tail. You get the real room's fingerprint with an adjustable decay. LiquidSonics' approach of interpolating between multiple measured IRs and adding modulation is a commercial expression of the same idea, and the research equivalent is fitting a differentiable FDN to a measured RIR, which is where the field is now.

Geometry, Rays And Wave Solvers

A separate discipline computes impulse responses from a room description instead of measuring or approximating one. It matters here for two reasons: it produces the early reflections that hybrid reverbs use, and it is what every game engine and VR audio renderer actually runs.

Image source method, 1979

Allen and Berkley, Image method for efficiently simulating small-room acoustics (JASA 65(4), 1979), is the workhorse. Reflect the source across each wall plane to create a mirror image; the reflected path from source to listener is geometrically identical to the direct path from the image to the listener. Recurse to get higher orders.

Practical notes that bite people: images must be checked for visibility in non-convex rooms; the method models specular reflection only, so a real room's scattering has to be grafted on separately; and frequency-dependent absorption means each image path needs its own filter, not a scalar gain. The method's real home is generating the first 50 ms, which is exactly the region where geometric acoustics is valid.

Ray and beam tracing, radiance transfer

Shoot rays (or cones, or beams) from the source, bounce them off surfaces with a specular/diffuse split, and histogram the energy that reaches the receiver as a function of time. Scales to arbitrary geometry, handles diffuse scattering naturally, and is embarrassingly parallel. Costs: stochastic noise in the resulting energy histogram, no phase information (so you get an energy decay curve rather than a true IR and must synthesise the fine structure with noise), and no diffraction. Acoustic radiance transfer precomputes surface-to-surface energy exchange, which lets you move the listener cheaply, at the cost of a precomputation stage.

Solving the wave equation directly

Below the Schroeder frequency, rays are wrong: wavelengths are comparable to room dimensions, modes dominate, and diffraction is the main effect. There you solve the PDE.

Scattering delay networks deserve a note because they close the loop back to FDN. De Sena, Hacıhabiboğlu, Cvetković and Smith (IEEE TASLP 23(9), 2015) place a scattering junction at the centre of each wall and connect junctions with bidirectional delay lines whose lengths come from the actual room geometry. First-order reflections are then exactly correct (they are the direct wall-to-wall paths), the late field emerges from recirculation as in an FDN, and the whole thing is parameterised by room dimensions and wall absorption rather than by tuned coefficients. It is the cleanest bridge between geometric accuracy and FDN efficiency, and it is what a modern interactive renderer wants.

The hybrid frequency split

      how a serious simulation splits the problem, both axes at once

            │  low frequency             │  high frequency
            │  below f_s (modal)         │  above f_s (statistical)
  ──────────┼────────────────────────────┼───────────────────────────
   early    │  FDTD / ARD / FEM          │  image source method
   t < t_mix│  diffraction matters,      │  specular, exact for
            │  rays are invalid          │  shoeboxes, cheap to
            │                            │  order 2 or 3
  ──────────┼────────────────────────────┼───────────────────────────
   late     │  modal synthesis           │  FDN / ray-traced EDC
   t > t_mix│  a few dozen resonators    │  + noise-shaped tail,
            │  for the strong modes      │  per-band T60 fitted
            │                            │  to the simulation

  Crossovers at f_s (from √(T60/V)) and at t_mix (from √V ms). Both
  numbers come straight out of section 02 and 03. Nothing in this
  field is arbitrary; it is all the same two boundaries.

Modern Practice · 2010 to now

The classical topologies are settled. What has changed in the last fifteen years is that the parameters no longer have to be tuned by ear, and that the FDN has been rediscovered as a general-purpose framework rather than one algorithm among several.

The FDN as a universal container

A useful result from the recent literature: most historical reverb topologies can be written as an FDN with a particular feedback matrix and particular filters. Schroeder's parallel combs are an FDN with a diagonal matrix. His series allpasses are an FDN with a triangular one. Moorer's extension, Dattorro's allpass feedback network and Dahl and Jot's absorbent allpass FDN all fall out of the same equations. That matters practically: if you implement one well-optimised FDN engine you can express the whole family in it, and you can reason about all of them with the same losslessness and decay theory.

Velvet noise

A late reverberation tail is perceptually equivalent to filtered noise, and you do not need a recursive network to make noise. Velvet noise is a sparse sequence of impulses of value +1 or -1, one per fixed-length interval at a randomised position within it, at a density of roughly 1,000 to 2,000 impulses per second. It sounds smoother than Gaussian white noise at the same density (hence the name) and, because it is sparse and its values are ±1, convolving with it needs only additions and subtractions: no multiplies.

Differentiable reverb

The current research front. Write the FDN in a framework that supports automatic differentiation, define a loss against a target (a measured room impulse response, or a spectral flatness objective), and optimise the delay lengths, feedback matrix and filter coefficients by gradient descent. Two distinct uses:

• Colourless reverberation. Dal Santo, Prawda, Schlecht and Välimäki (DAFx 2023) optimise a differentiable FDN to minimise spectral coloration, using a spectral loss for flat magnitude plus a temporal loss to control impulse-response sparsity. The finding is that even very small FDNs produce high-quality colourless reverb once optimised, which overturns the assumption that you need large N.
• Fitting measured rooms. Mezza, Giampiccolo, De Sena and Bernardini fit differentiable FDNs with learnable delay lines to measured RIRs (EURASIP JASMP 2024), extend it to MIMO for multichannel RIRs and higher-order microphone arrays (DAFx 2024, DAFx 2025), and Differentiable Scattering Delay Networks (DAFx 2025) do the same for geometry-parameterised SDNs. Differentiable Grouped FDNs (2025) learn a mapping from source and listener position to network gains via a small MLP, which gives real-time adaptation to movement at a fraction of the cost of modal methods.

Supporting infrastructure exists: Schlecht's FDNTB (the Feedback Delay Network Toolbox, DAFx 2020) collects the special matrices, topologies, attenuation filters and modal decompositions; FLAMO (ICASSP 2025) is an open-source library for frequency-domain differentiable audio processing with learnable LTI modules. Lee, Choi and Lee's Differentiable Artificial Reverberation (IEEE/ACM TASLP 30, 2022) is the general framing paper.

Beyond a single exponential

Sabine's model assumes one diffuse field with one decay rate per band. Real spaces frequently violate this: coupled volumes (a nave and its side chapels, a stage house and an auditorium, a stairwell off a corridor) produce a sum of exponentials, so the energy decay curve is bent rather than straight on a dB axis. The common slopes model represents a space as a small set of shared decay rates with position-dependent amplitudes, which is a good fit for coupled rooms and is what several current XR renderers use. Neural approaches to multi-exponential decay analysis (Götz, Pérez, Schlecht, Pulkki, JASA 152(2), 2022) fit these automatically.

The plugin landscape, honestly

The shimmer family is worth a paragraph because it is the clearest case of a reverb topology used as a compositional device. Put a pitch shifter inside the feedback loop (typically +12 semitones, sometimes +12 and -12 in parallel) and every circulation transposes the tail upward, producing an ascending harmonic cloud that never resolves. It comes out of Brian Eno and Daniel Lanois's work in the 1980s, notably on U2, the original chain is usually reconstructed as an AMS DMX 15-80's pitch shifter feeding a Lexicon 224 on Concert Hall (Eventide harmonisers belong to later recreations of the trick). It is now a preset, but its mechanism is exactly the loop from Allpass Loops with one nonlinear block inserted, and it is a useful reminder that the topology does not care whether the block is physical.

Build It · Reference Implementations

Working code, tested. NumPy sample-by-sample loops for clarity rather than speed: the point is that every structure on this page is fifty lines. Ship the vectorised version later.

Primitives and decay

pythonimport numpy as np

def comb_gain(m, T60, sr):
    """Feedback gain giving a -60 dB decay in T60 seconds for an m-sample loop."""
    return 10.0 ** (-3.0 * m / (sr * T60))

class Comb:
    # Moorer's lowpass comb: damp=0 is Schroeder's plain comb
    def __init__(self, m, g, damp=0.0):
        self.buf = np.zeros(m); self.i = 0
        self.g = g; self.damp = damp; self.lp = 0.0
    def tick(self, x):
        y = self.buf[self.i]
        self.lp = y * (1.0 - self.damp) + self.lp * self.damp   # one-pole LPF
        self.buf[self.i] = x + self.g * self.lp
        self.i = (self.i + 1) % len(self.buf)
        return y

class Allpass:
    """Dattorro / Gardner two-multiply form. One buffer, exactly allpass."""
    def __init__(self, m, g):
        self.buf = np.zeros(m); self.i = 0; self.g = g
    def tick(self, x):
        d = self.buf[self.i]
        t = x - self.g * d
        self.buf[self.i] = t
        self.i = (self.i + 1) % len(self.buf)
        return d + self.g * t

Schroeder / Moorer

pythondef schroeder_moorer(x, sr=48000, T60=2.0, damp=0.4):
    cm = [1687, 1601, 2053, 2251]          # mutually prime comb delays
    am = [347, 113]                        # short prime allpass delays
    combs = [Comb(m, comb_gain(m, T60, sr), damp) for m in cm]
    aps   = [Allpass(m, 0.7) for m in am]
    y = np.zeros(len(x))
    for n, s in enumerate(x):
        v = sum(c.tick(s) for c in combs) / len(combs)
        for a in aps:
            v = a.tick(v)
        y[n] = v
    return y

FDN with unitary feedback and Jot damping

pythondef hadamard(n):
    H = np.array([[1.0]])
    while H.shape[0] < n:
        H = np.block([[H, H], [H, -H]])
    return H / np.sqrt(n)                     # orthonormal: H.T @ H == I

def fdn(x, sr=48000, T60=2.0, m=(997, 1153, 1327, 1559), hf_ratio=0.35):
    """N-line FDN, unitary feedback, per-line absorbent shelf (Jot 1991)."""
    m = np.asarray(m, dtype=int); N = len(m)
    A = hadamard(N)
    g_dc = 10.0 ** (-3.0 * m / (sr * T60))              # gain at DC
    g_hf = 10.0 ** (-3.0 * m / (sr * T60 * hf_ratio))   # gain at Nyquist
    a  = (g_dc - g_hf) / (g_dc + g_hf)                  # one-zero shelf: H = b0*(1 + a*z^-1)
    b0 = (g_dc + g_hf) / 2.0
    bufs = [np.zeros(mi) for mi in m]
    idx  = np.zeros(N, dtype=int)
    z    = np.zeros(N)                                   # shelf state
    bvec = np.ones(N) / np.sqrt(N)                       # input gains
    cvec = np.ones(N) / np.sqrt(N)                       # output gains
    y = np.zeros(len(x))
    for n, s in enumerate(x):
        out  = np.array([bufs[i][idx[i]] for i in range(N)])
        y[n] = cvec @ out
        filt = b0 * (out + a * z); z = out
        fb   = A @ filt + bvec * s
        for i in range(N):
            bufs[i][idx[i]] = fb[i]
            idx[i] = (idx[i] + 1) % m[i]
    return y

Measurement, so you can audit your own output

pythondef t30(h, sr):
    """T60 estimate: Schroeder backward integration, least-squares line fit
    over the -5 to -35 dB span of the EDC, extrapolated to -60 dB."""
    e  = np.cumsum(h[::-1] ** 2)[::-1]
    db = 10 * np.log10(e / e[0] + 1e-30)
    i5, i35 = np.argmax(db <= -5), np.argmax(db <= -35)
    n = np.arange(i5, i35 + 1)
    slope = np.polyfit(n / sr, db[i5:i35 + 1], 1)[0]   # dB per second
    return -60.0 / slope

def echo_density(h, sr, half_win_ms=20):
    """Abel-Huang normalised echo density. 1.0 means Gaussian-noise-like.
    Window spans n +/- half_win_ms (40 ms total at the default)."""
    w   = max(1, int(sr * half_win_ms / 1000.0))
    ref = 0.3173                       # erfc(1/sqrt(2)), the Gaussian value
    out = np.zeros(len(h))
    for n in range(w, len(h) - w):
        seg = h[n - w:n + w]
        sd  = seg.std()
        out[n] = (np.abs(seg) > sd).mean() / ref if sd > 0 else 0.0
    return out

What the measurements say

Run at 48 kHz with a unit impulse, T60 target 2.0 s. All numbers below are measured from the code above, not estimated. The input diffusers are Dattorro's published set (allpass delays 142, 107, 379, 277 with g = 0.75, 0.75, 0.625, 0.625); the N = 8 delay set is [997, 1153, 1327, 1559, 1801, 2099, 2311, 2521]. Octave-band values were measured through a cascaded biquad bandpass.

Adding what the classics added

1. Pre-delayA plain delay line in front of everything, 0 to 150 ms. The single most effective control for separating a source from its space, and the one that sets apparent room size (see ITDG in IR Anatomy).
2. Explicit early reflectionsA sparse tapped delay line, 12 to 20 taps, with times and gains from the image source method for the room you want. Sum it with the FDN output under an independent level control, exactly as Moorer proposed in 1979 and TC Electronic productised in VSS.
3. ModulationSlowly vary the delay read positions with independent low-frequency noise per line, depth of a few samples, rate under 1 Hz. Use allpass or Lagrange fractional-delay interpolation, not truncation, or you will hear the modulation as noise. This buys smoothness and stability margin at the cost of phase coherence.
4. Tonal correctionAn output shelf inversely proportional to T60(ω), so that shortening the high-frequency decay does not also make the reverb dull.
5. Input and output diffusionSeries allpasses before the network (as measured above) and optionally after it. Cheapest possible improvement in perceived quality.
6. Low cut and high cut on the wet pathNot physics, but every engineer does it: rolling off below 150 Hz and above 8 kHz on the reverb return is how you get a big space that does not fight the mix.

Further · Papers, Errata, Rabbit Holes

The foundational ten

If you read only these, in this order, you will know the field. The list follows Sean Costello's (Valhalla DSP) selection, which is the best curated bibliography in the field and worth reading in his own words.

1. Schroeder & Logan, "Colorless" Artificial Reverberation, JAES 9(3), July 1961. The allpass delay. PDF
2. Schroeder, Natural Sounding Artificial Reverberation, JAES July 1962. Parallel combs plus series allpass, and the matrix transpose footnote that seeded FDNs. AES e-lib
3. Gerzon, Synthetic Stereo Reverberation Part 1, Studio Sound December 1971, pp. 632-635. Orthogonal-matrix feedback networks, ten years early. Scan
4. Gerzon, Synthetic Studio Reverberation Part 2, Studio Sound January 1972, pp. 24-28. Allpass feedback networks, cascaded unitary networks, filters inside unitary structures. Scan
5. Moorer, About This Reverberation Business, Computer Music Journal 1979. Lowpass combs, the efficient allpass, explicit early reflections. Full text at IRCAM
6. Stautner & Puckette, Designing Multi-Channel Reverberators, CMJ 6(1), Spring 1982. The FDN, by that name, with code. PDF
7. Smith, A New Approach to Digital Reverberation using Closed Waveguide Networks, ICMC 1985. Waveguides and scattering junctions. PDF
8. Jot & Chaigne, Digital delay networks for designing artificial reverberators, AES 90th Convention, Paris, February 1991. Per-delay absorbent filters, uniform decay, tonal correction. AES e-lib
9. Gardner, A Realtime Multichannel Room Simulator, ASA New Orleans, November 1992. Allpass loops, nested allpass, single-buffer implementation. PDF
10. Dattorro, Effect Design Part 1: Reverberator and Other Filters, JAES 45(9), September 1997. A complete plate reverb with every coefficient, plus Blesser's commentary on why modulation exists. PDF

Then these

• Gerzon, Unitary (energy preserving) multi-channel networks with feedback, Electronics Letters 12(11), 1976. The theory made formal.
• Rocchesso & Smith, Circulant and elliptic feedback delay networks for artificial reverberation, IEEE Trans. Speech & Audio 5(1), 1997. Feedback matrices you can apply with an FFT, which is how large N becomes affordable. (Frequently miscited as "Rochesso" or attributed to a nonexistent "Joël Rochesso"; the author is Davide Rocchesso.)
• Jot, An analysis/synthesis approach to real-time artificial reverberation, ICASSP 1992, and his 1992 Telecom Paris doctoral thesis. The perceptual parameterisation behind Spat and a generation of spatialisers.
• Allen & Berkley, Image method for efficiently simulating small-room acoustics, JASA 65(4), 1979. Where early reflections come from.
• Schroeder, New Method of Measuring Reverberation Time, JASA 37, 1965, and Integrated-impulse method measuring sound decay without using impulses, JASA 66(2), 1979. Backward integration, and the energy decay curve.
• Farina, Simultaneous measurement of impulse response and distortion with a swept-sine technique, AES 108th Convention, 2000. The exponential sine sweep.
• Välimäki, Parker, Savioja, Smith, Abel, Fifty Years of Artificial Reverberation, IEEE TASLP 20(5), 2012, and the later More Than 50 Years of Artificial Reverberation. The two survey papers; start here if you want one document instead of ten.
• Schlecht & Habets, On Lossless Feedback Delay Networks, IEEE TSP 65(6), 2017, and Schlecht's thesis Feedback Delay Networks in Artificial Reverberation and Reverberation Enhancement. Current FDN theory.
• De Sena, Hacıhabiboğlu, Cvetković, Smith, Efficient synthesis of room acoustics via scattering delay networks, IEEE/ACM TASLP 23(9), 2015.
• Välimäki, Holm-Rasmussen, Alary, Lehtonen, Late reverberation synthesis using filtered velvet noise, Applied Sciences 7(5), 2017.
• Dal Santo, Prawda, Schlecht, Välimäki, Differentiable feedback delay network for colorless reverberation, DAFx 2023. And the 2024/2025 Mezza and Giampiccolo papers on differentiable FDNs and SDNs.
• Blesser, Spaces Speak, Are You Listening? Experiencing Aural Architecture, MIT Press, 2007. By the designer of the EMT 250's algorithms, and not a DSP book: it is about what reverberation means culturally and perceptually. Read it last, or first.
• Physical modelling of the mechanical devices: Bilbao and Parker on spring reverberation (numerical and modal), Välimäki, Parker and Abel on dispersion-generation structures for spring emulation, and Abel and Smith's patents on dispersive-delay and modal-decomposition reverberation.
• Sean Costello's Valhalla DSP blog. Decades of design notes from a working reverb developer, including the Ursa Major analysis and the paper list above.
• Julius O. Smith's online books at ccrma.stanford.edu, particularly Physical Audio Signal Processing, which contains the canonical FDN history and derivations, free and hyperlinked.

Errata · claims that circulate and do not hold up

Collected from checking the popular account against primary and contemporary sources. If you have a reverb explainer in your notes, it probably contains several of these.

Rabbit holes worth the time

• Plot the poles. Take a lowpass comb filter, plot pole locations as g and d vary, and watch the decay-time-versus-frequency relationship appear geometrically. Then do it for an N=4 FDN and see Σmᵢ poles arranged on a near-circle whose radius varies with frequency. This makes the Jot damping argument obvious in a way no formula does.
• Implement Dattorro exactly. All the coefficients are published. Do it at the paper's 29,761 Hz, then again at 48 kHz with rescaled and re-primed delays, and listen to the difference. It is a lesson in why "the same algorithm" is not the same sound.
• Measure something. A laptop, an interface, one speaker, one microphone and a 20-second exponential sweep will give you a usable IR of your stairwell. Deconvolve it yourself rather than using a tool, crop the negative-time distortion products, and you will understand Farina's paper properly.
• Audit a commercial reverb. Feed a plugin an impulse, capture the output, and run the echo density and per-band T30 code from Build It on it. You can usually tell whether it is an FDN, an allpass loop or a convolution, whether the modulation is periodic or random, and whether the decay control is calibrated.
• Build a spring emulator. Cascade forty first-order allpasses, tune the coefficients so group delay rises with frequency toward a cutoff in the few-kilohertz region, and put the chain in a feedback loop. You will get a drip. Then read Bilbao and Parker and find out how much you were missing.
• Fit an FDN by gradient descent. FLAMO and FDNTB are both public. Take a measured IR, fit a small differentiable FDN to its energy decay relief, and compare against your hand-tuned version. This is a weekend, and it is the current state of the art.