Architectural Foundations of Infinite Auditory Memory: Recursive Temporal Networks and Non-Linear Feedback Systems

Full report

On this page

The conceptualization of an auditory equivalent to visual mesmerization—an endlessly receding, self-transforming cascade of sensory memory—requires transcending traditional paradigms of spatial reverberation. In visual media, specifically the recursive spatial effects found in video feedback loops (such as those exploring spatial infinities on platforms like mesmerization.com), a frame buffer feeds back into itself while undergoing continuous spatial, chromatic, and opacity transformations1. In the auditory domain, the equivalent phenomenon relies on highly optimized temporal-memory systems. When digital delay loops, feedback networks, and recursive spectral transformations are mathematically unbounded yet systemically stabilized, sound ceases to act merely as an acoustic reflection. Instead, it functions as a living memory. New acoustic events continuously collide with, and transform, the remanent decay of earlier sounds, producing an infinitely morphing auditory topology.

To achieve this state of "auditory infinity," audio engineers must deploy highly sophisticated Digital Signal Processing (DSP) architectures. The design necessitates a transition from elementary multi-tap delays to unitary Feedback Delay Networks (FDNs), introducing non-linear saturation via Antiderivative Antialiasing (ADAA), and establishing a rigorous algorithmic policy—defined herein as a TemporalAudioPolicy—to safeguard against mathematical collapse. This report provides an exhaustive analysis of the required signal processing methodologies, mathematical stability proofs, and software architecture necessary to render endless, self-transforming sound.

I. Topologies of Delay: From Discrete Echoes to Infinite Diffusion

The sensation of acoustic memory begins with the delay line, fundamentally represented in the z-domain by the discrete-time operator [Figure omitted from source export], where [Figure omitted from source export] is the delay length in samples. Traditional delay architectures route discrete remanence across the stereo field, but they fail to produce the exponentially dense, unified wash of sound required for true auditory mesmerization.

Elementary Delay Structures

At the foundational level, digital delay lines capture an audio frame buffer and release it after a specified temporal distance. A basic feedback delay routes a percentage of the output back into the input, yielding a decaying train of discrete echoes. Multi-tap delays extend this by reading from multiple equidistant or logarithmically spaced points within the same delay buffer, simulating the early specular reflections of a physical space. Ping-pong delays cross-wire the feedback paths between the left and right stereo channels, creating a localized spatial widening effect.

While useful for rhythmic echoing, these elementary topologies suffer from periodicity. The echoes remain distinct, failing to blur into a continuous memory stream. To transform discrete echoes into a continuous tail, filtered feedback and diffusion must be introduced. By embedding a lowpass filter within the feedback loop, subsequent echoes lose high-frequency energy, simulating the frequency-dependent air absorption of natural spaces and pushing the sound perceptually backward in depth. Diffusion, achieved by scattering the phase of the signal, begins to break the distinct transients into a continuous wash.

The Schroeder and Moorer Architectures

The earliest mathematical formalization of infinite, colorless diffusion was proposed by Manfred Schroeder and B. Logan in 19614. Schroeder’s reverberator sought to build infinite recursive density using a parallel bank of feedback comb filters fed into a series of allpass networks. The Schroeder allpass filter utilizes a feedforward and feedback loop that preserves the signal’s energy across all frequencies, altering only the phase6. Because the magnitude response remains strictly flat, an allpass network can recursively process a signal infinitely without inducing runaway resonance at any specific frequency.

James A. Moorer subsequently advanced the Schroeder paradigm in 1979 by substituting Schroeder’s standard comb filters with lowpass feedback comb filters9. Moorer's transfer function allowed for frequency-dependent decay, simulating the high-frequency air absorption of natural acoustic spaces more accurately than Schroeder's original model9.

The Dattorro Figure-Eight Tank and Cross-Feedback

In 1997, Jon Dattorro published a highly influential reverberation topology—the "tank" structure—which remains a cornerstone of algorithmic effect design11. Dattorro's reverberator explicitly isolates the initial early reflections (managed via pre-delay and tapped discrete delays) from the late reverberation tail13.

The tank itself operates as a giant figure-eight cross-feedback loop. Inside the tank, the signal passes through cascaded, modulated allpass filters, absorptive lowpass filters, and massive delay lines, crossing over into the opposite channel iteratively12. The Dattorro architecture demonstrates how cross-feedback between left and right delay networks creates immense echo density and phase decorrelation12. By feeding the remanence of the left channel into the right, and vice versa, the system generates a psychological illusion of endless, surrounding space, ensuring that the auditory memory never stagnates in a localized stereo position.

II. Feedback-Delay Networks (FDNs) and Matrix Mathematics

While the Schroeder and Dattorro models achieve diffuse persistence, the Feedback Delay Network (FDN), popularized by Jean-Marc Jot and Antoine Chaigne in 1991, provides the generalized mathematical framework required for absolute control over infinite auditory memory4. An FDN effectively vectorizes the comb filter concept, allowing for highly complex, multi-dimensional signal scattering.

System Formulation and the Generalized Characteristic Polynomial

An FDN consists of [Figure omitted from source export] parallel delay lines whose outputs are fed back into their inputs via an [Figure omitted from source export] feedback matrix [Figure omitted from source export]. The discrete-time transfer function of a single-input, single-output (SISO) FDN is expressed as:

[Figure omitted from source export]

where [Figure omitted from source export] is the delay matrix, [Figure omitted from source export] and [Figure omitted from source export] are the input and output gain vectors, and [Figure omitted from source export] is the direct feedthrough gain4.

The poles of the FDN—which dictate the resonant frequencies and the infinite stability of the system—are the roots of the Generalized Characteristic Polynomial (GCP):

[Figure omitted from source export]

1. For the system to remain stable—meaning the auditory memory does not run away into infinite clipping—all roots of the GCP must lie strictly inside the unit circle of the complex plane ([Figure omitted from source export])4.

To achieve a "freeze-like" near-unity feedback state, where the auditory memory sustains indefinitely without decaying or exploding, the roots must lie exactly on the unit circle. This is achieved when the feedback matrix [Figure omitted from source export] is lossless. Early FDNs utilized strictly unitary matrices ([Figure omitted from source export]), such as the Hadamard or Householder matrices19. More recently, research has proven that any matrix that is diagonally similar to a unitary matrix (unilossless) guarantees that the FDN is energy-preserving regardless of the delay lengths chosen19.

Modal Decomposition and the Eradication of Coloration

A primary flaw in recursive audio is "coloration"—the emergence of metallic ringing caused by the non-uniform excitation of resonant modes. When sound acts as an infinite memory, any slight resonant peak will accumulate over thousands of feedback cycles, eventually dominating the signal with a piercing frequency.

Schlecht and Habets introduced methods to analyze FDNs via modal decomposition, representing the FDN's impulse response as a sum of complex one-pole resonators18. By utilizing the Ehrlich-Aberth iteration on the GCP, DSP engineers can accurately map the pole frequencies and decay rates of massive networks18. To create a mesmerizing, colorless infinite wash, the FDN's parameters can be optimized using stochastic gradient descent. By penalizing sparse parameter values and minimizing spectral loss, a dense, flat modal excitation is achieved, eliminating metallic artifacts even when the memory sustains indefinitely4.

Filter Feedback Matrices (FFM) and Uniallpass Networks

To maximize the density of the auditory memory—transforming discrete echoes into a smooth continuum almost instantaneously—the scalar matrix [Figure omitted from source export] can be replaced with a Filter Feedback Matrix (FFM), where each matrix entry is an FIR or IIR filter rather than a simple gain multiplier25.

A specialized variant of the FFM, known as the "velvet feedback matrix," creates ultra-dense reflections resembling non-specular acoustic scattering (rough boundary diffusion) at a minimal computational cost25. Furthermore, a "uniallpass" FDN preserves the allpass characteristic across all frequencies for any arbitrary choice of delay lengths, establishing the ultimate colorless infinite tail6. By combining uniallpass structures with FFMs, the auditory memory achieves total diffusion, erasing the original transients of the input signal and leaving only its spectral ghost.

III. Modulated FDNs and Orthogonal Matrix Parameterization

If an auditory memory is to be self-transforming, its spatial and geometric parameters must evolve over time. However, exposing the feedback matrix [Figure omitted from source export] to real-time user modulation poses a severe stability risk. If the matrix coefficients are arbitrarily interpolated, the matrix may temporarily break the unitary condition, pushing eigenvalues outside the unit circle and causing immediate runaway feedback.

Parameterization via Givens Rotations

To parameterize an orthogonal matrix such that it can be safely and continuously modulated, Givens rotations are utilized30. A Givens rotation matrix [Figure omitted from source export] performs a two-dimensional rotation by an angle [Figure omitted from source export] in the [Figure omitted from source export] plane of an [Figure omitted from source export]\-dimensional space32. The product [Figure omitted from source export] represents a counterclockwise rotation of the vector [Figure omitted from source export]32.

Any [Figure omitted from source export] orthogonal matrix can be decomposed into a product of [Figure omitted from source export] Givens rotations31. By attaching low-frequency oscillators (LFOs) or macroscopic user controls to the angles ([Figure omitted from source export]) of these constituent Givens rotations, the FDN’s internal scattering geometry twists and morphs in real-time. The resulting matrix remains strictly orthogonal (lossless) at every sample31. This allows the sound to recirculate endlessly, shifting its spatial and phase relationships, panning organically across the psychoacoustic field, without ever accumulating runaway gain. This satisfies the psychological requirement for mesmerization: the sonic environment is constantly moving, yet mathematically contained.

IV. Fractional Delays and Time-Varying Memory

Auditory mesmerization relies heavily on micro-pitch modulation within the reverberant tail. If the delay lengths [Figure omitted from source export] are entirely static, the memory feels rigid and artificial. If [Figure omitted from source export] fluctuates smoothly, the Doppler effect causes the recirculating sound to pitch-shift, chorus, and smear, creating a lush, organic texture characteristic of high-end vintage hardware and natural fluid acoustic environments33.

However, altering a digital delay line's length dynamically requires sub-sample precision. Simply truncating or rounding to the nearest integer sample causes catastrophic audio artifacts known as zipper noise. Fractional delay filters are therefore strictly required to interpolate the spaces between discrete digital samples. While Lagrange interpolation (FIR) is commonly used for standard delay effects, it inherently attenuates high frequencies during interpolation, which degrades an infinite feedback loop over time34.

Thiran Allpass Interpolation

For lossless, infinite memory, fractional delay is best implemented using Thiran allpass filters. The Thiran allpass is a recursive IIR filter designed to possess a maximally flat group delay at low frequencies (near DC)36. Because its magnitude response is perfectly flat (allpass), it seamlessly approximates fractional time shifts without dampening high frequencies during infinite recirculation.

The coefficients [Figure omitted from source export] of an [Figure omitted from source export]\-th order Thiran allpass filter for a continuous fractional delay [Figure omitted from source export] are derived via closed-form combinatorial formulas:

[Figure omitted from source export]

1. While providing excellent phase accuracy, the Thiran allpass becomes unstable if the continuous delay parameter [Figure omitted from source export] drops below [Figure omitted from source export]36. By truncating the Thiran coefficient vector or strictly enforcing a baseline integer delay within the delay line loop, unconditional stability is guaranteed35. Embedding time-varying Thiran allpass filters inside the FDN ensures that the auditory memory continuously detunes and drifts, introducing a rich jitter density that mimics the chaotic variations of analog magnetic tape.

V. Recursive Spectral Transformations

To create sound that is truly "endlessly receding and self-transforming," the feedback loop must contain active spectral modifiers39. Rather than acting as a static simulated room, the FDN acts as a generative synthesis engine, tearing apart and rebuilding the audio on every cycle.

Pitch-Shifting Feedback

Injecting pitch-shifters directly into the feedback loop causes each recursive pass to shift upward or downward in frequency, creating Shepard-tone-like infinite ascensions or ethereal "shimmer" reverbs40. Implementing pitch shifting inside a micro-delay loop is highly complex. Standard phase vocoders, which rely on Fast Fourier Transforms (FFTs) and bin-wise phase unwrapping, introduce significant block latency43. This latency disrupts the tight temporal timing of the FDN and ruins the density of the reverb tail. Therefore, time-domain granular pitch shifting or modulated variable-delay-line resampling must be utilized to maintain zero or near-zero latency within the feedback path41.

Granular Feedback and Jitter Density

Granular feedback introduces a profound layer of temporal breakdown, directly analogous to the pixel-sorting and visual tearing seen in visual mesmerization loops. By placing a granular processor—which slices the audio stream into 10–50ms grains, applies a Hann or Gaussian amplitude envelope, and re-pitches, reverses, or scrambles them—inside the FDN loop, the echoes do not merely fade42. They disintegrate.

This process creates "jitter density," physically breaking the deterministic periodicity of the FDN. As grains overlap and feed back into the orthogonal scattering matrix, the sound dissolves into a cloud of microsound, yielding a truly chaotic, non-repeating auditory infinity where the original source material is entirely subsumed by its own granular memory42.

VI. Frequency-Selective Persistence and the "Memory" Control

A public "Memory" or "Echo" macro-control must act in a musically meaningful way across all synthesis modes, scaling seamlessly from a dry transient to a cavernous infinite wash. This is executed by mapping the user-facing Memory parameter to the frequency-dependent reverberation time ([Figure omitted from source export]) of the network.

To achieve frequency-selective persistence—where low frequencies might rumble indefinitely while high frequencies evaporate, or vice versa—absorptive shelf or biquad filters are embedded serially within the delay lines8. The required target loop gain [Figure omitted from source export] for delay line [Figure omitted from source export] at frequency [Figure omitted from source export] is mathematically derived from the delay length [Figure omitted from source export] and the desired [Figure omitted from source export]:

[Figure omitted from source export]

where [Figure omitted from source export] is the system sampling rate47. By exposing the [Figure omitted from source export] contour to the user, the infinite memory can be sculpted. A low-pass decay curve mimics physical reality (air absorption), while a high-pass decay curve creates an unnatural, ethereal memory that dissolves into a whispering hiss. When the Memory control is pushed to 100%, the filters are bypassed, forcing [Figure omitted from source export], achieving the theoretical freeze state.

VII. Non-Linear Saturation and Antiderivative Antialiasing (ADAA)

If an infinite memory merely stacks linearly, it eventually becomes a mathematically flat, sterile wall of white noise. To mimic the psychological mesmerization of a vivid, organic memory—similar to the analog degradation of degraded VHS tape or tube amplification—non-linear saturation must be applied inside the FDN loop39.

However, inserting a basic non-linear waveshaper, such as [Figure omitted from source export], into a recursive discrete-time loop creates two catastrophic DSP failures:

1. Unbounded Energy (Loss of Passivity): The non-linearity may inadvertently inject energy into the system, pushing the loop gain above unity and causing the FDN to explode into numeric overflow.

2. Aliasing: Non-linearities generate upper harmonics that exceed the Nyquist frequency limit. In a feedback loop, these aliased frequencies fold back into the audible spectrum, intermodulate, and compound upon every single cycle, quickly destroying the sound with harsh, atonal digital artifacts49.

Port-Hamiltonian Bounds and Strict Passivity

To ensure the saturator does not cause runaway feedback, it must be proven strictly passive. Advanced virtual analog modeling relies on Port-Hamiltonian system theory or Wave Digital Filters (WDF) to ensure that the non-linearity is modeled such that the energy leaving the node never exceeds the energy entering it49. If the FDN is configured as a lossless scattering junction, and the non-linearity is strictly bounded (e.g., [Figure omitted from source export] for all [Figure omitted from source export]), the global system remains strictly contractive, absolutely preventing infinite gain explosions even at infinite memory settings52.

Antiderivative Antialiasing (ADAA)

To solve the aliasing problem without resorting to massive CPU-intensive oversampling of the entire FDN architecture, Antiderivative Antialiasing (ADAA) is employed49. ADAA analytically approximates the continuous-time convolution of the non-linear function with a bandlimiting kernel before sampling it back into the digital domain.

For a memoryless non-linearity [Figure omitted from source export], the first-order ADAA operator evaluates the analytical antiderivative [Figure omitted from source export]. The discrete-time antialiased output [Figure omitted from source export] is computed using the finite difference:

[Figure omitted from source export]

1. Because ADAA operates on the difference between the current and previous sample, it acts as a localized low-pass filter and inherently introduces a continuous fractional delay of approximately half a sample49. In an FDN, this fractional delay must be precisely subtracted from the corresponding delay line [Figure omitted from source export] to maintain exact phase alignment and the uniallpass condition49. By embedding ADAA [Figure omitted from source export] saturators at the nodes of the Givens-rotated feedback matrix, the auditory memory gracefully degrades, saturates, and compresses, mimicking the rich harmonic density of analog mesmerization without any destructive aliased artifacts39.

VIII. Establishing Stability: Mathematical Conditions and Engineering Guards

Creating a software system capable of running an endlessly receding memory for hours—such as in a live interactive installation, a generative music stream, or a prolonged studio session—requires absolute systemic safety. Mathematical drift over billions of samples will inevitably find edge cases.

The following engineering guards establish the mathematical conditions preventing system collapse:

 

IX. TemporalAudioPolicy, Graph Architecture, and Testing

To translate this complex mathematics into a reusable, production-ready C++ plugin architecture (applicable across VST3, AU, and AAX formats), the engine requires a definitive TemporalAudioPolicy. This policy dictates the lifecycle, memory management, and restart/stop semantics of the recursive sound, bridging the gap between theoretical DSP and host DAW (Digital Audio Workstation) interactions60.

Directed Acyclic Graph (DAG) Architecture

The software architecture is structured as a Directed Acyclic Graph (DAG) for the feedforward components, containing an explicitly managed internal cyclical sub-graph representing the FDN49.

1. Source Node: Handles dry audio input and input gain scaling.

2. Pre-Conditioning Node: Injects early reflection multi-taps, initial diffusion allpasses, and mild ADAA saturation.

3. Memory Core (FDN Node): The massive recursive engine containing the unitary matrix, circular delay buffers, Thiran allpasses, and [Figure omitted from source export] absorptive filters.

4. Spectral Transform Node: Contains granular processors and pitch shifters, sitting inline with the delay buffers inside the FDN loop.

5. Master Output Node: Handles Dry/Wet mix blending and global DC blocking.

Restart/Stop Semantics and Tail Processing

A critical aesthetic and functional issue in infinite reverberation is the behavior of the DAW when the transport is stopped. If the system stops rendering audio to the speakers, but the plugin's internal delay buffers remain populated, pressing "play" again will instantly blast the user with the loud remnants of the previous infinite memory (a "ghost echo").

The TemporalAudioPolicy state machine enforces the following strict semantics:

• ACTIVE: Normal processing. Input is fed into the FDN; output is continuously generated.
• TAIL\_DECAY: Triggered when input falls to absolute silence. Input is muted (0.0), but the FDN loop continues processing. This state is communicated to the host via the AudioProcessor tail length parameter60. If the user has set the memory to "Infinite," the plugin actively overrides the host DAW's sleep/suspend commands to keep the tail alive.
• FROZEN: Input is muted. FDN loop gain is hard-locked to exactly 1.0. Absorption filters are bypassed. The current audio buffer cycles endlessly, morphing via Givens rotations without gaining or losing a single decibel of energy.
• FLUSHED (On Transport Stop/Restart): A strict policy where a clearBuffers() routine is mandated upon receiving a DAW transport stop flag. A std::fill command overwrites all circular delay lines, Thiran state variables, and ADAA previous-sample memories with exact 0.0f59. This strictly prevents source/node leaks and extremely long memory remnants persisting inappropriately after a stop61.

Reference and Long-Duration Stability Testing

To validate the TemporalAudioPolicy, exhaustive DSP unit testing is required:

1. Dirac Delta Injection Test: A single impulse is fed into the system with Memory set to 100%. The output is tracked for 10 million samples to ensure the energy envelope remains perfectly flat (no decay, no explosion), verifying the unilossless property of the Givens rotations.

2. Pink Noise Saturation Test: 0 dBFS pink noise is fed continuously into the ADAA saturators. The output spectrum is monitored via FFT to confirm the absolute absence of harmonic aliasing above the Nyquist limit.

3. Long-Duration Drift Test: The plugin is run continuously for 48 hours with a live input feed and random LFO modulations applied to all Thiran delays and Givens angles. A watchdog timer monitors for NaNs, denormals, or DC offsets exceeding 0.01. Passing this test certifies the engine for permanent live installation use.

X. Perceptual Presets: From Dry Immediacy to Auditory Infinity

To ensure that the mathematics serve the aesthetic goal of mesmerization, the highly complex underlying parameters (Givens angles, matrix density, T60 absorption, ADAA drive, Thiran bounds) are mapped to a high-level perceptual preset framework. The overarching control is the public "Memory" macro-parameter, which interpolates safely across synthesis modes.

XI. Conclusion

The pursuit of an auditory equivalent to visual mesmerization requires crossing the boundary between traditional spatial simulation and cognitive sound synthesis. By uniting the theoretical elegance of Feedback Delay Networks (FDNs) with the safely parameterizable stability of Givens orthogonal matrices, acoustic energy can be captured and fractured infinitely without triggering catastrophic loop failures. When fractional Thiran delays, ADAA-stabilized non-linear saturation, and granular pitch-shifting are folded into this strictly passive recursive loop, the echoes cease to merely repeat—they evolve autonomously.

Governed by a rigorous C++ TemporalAudioPolicy that ruthlessly manages denormal flushing, DC accumulation, and strict host stop/flush semantics, this architecture transcends the fragility of traditional audio feedback. It provides a robust, mathematically guaranteed engine for generating self-transforming, endlessly receding sound, achieving a true, living auditory memory.

Works cited

1. Clinical Psychology Dissertations \- Pacifica Graduate Institute, https://library.pacifica.edu/clinpsycdiss.html

2. ETD Template \- Digital Greensboro, https://www.digitalgreensboro.org/nanna/record/16897/files/Hunsberger\_uncg\_0154D\_14025.pdf?withWatermark=0\&withMetadata=0®isterDownload=1\&version=1

3. sonified biofeedback as generative art practice and theory, https://repository.londonmet.ac.uk/7598/1/536734.pdf

4. Feedback Delay Network Optimization \- arXiv, https://arxiv.org/html/2402.11216v1

5. Perceptual Similarities between Artificial Reverberation Algorithms, https://eprints.whiterose.ac.uk/id/eprint/225380/1/applsci-13-00840-v2.pdf

6. Allpass Feedback Delay Networks \- arXiv, https://arxiv.org/html/2007.07337v2

7. Frequency-Dependent Schroeder Allpass Filters \- MDPI, https://www.mdpi.com/2076-3417/10/1/187

8. Artificial Reverberation | Physical Audio Signal Processing, https://www.dsprelated.com/freebooks/pasp/Artificial\_Reverberation.html

9. Comparative Study of Optimal Design Strategies of Reverberators, https://benthamopen.com/contents/pdf/TOACOJ/TOACOJ-2-31.pdf

10. Shroeder-Moorer filter | RecordingBlogs, https://www.recordingblogs.com/wiki/shroeder-moorer-filter

11. Part Effect Design\* 1: Reverberator and Other Filters, https://www.semanticscholar.org/paper/Part-Effect-Design\*-1%3A-Reverberator-and-Other-Dattorro/7bff0e33b355afc9ae86118efa91f5ab4bf14746

12. Dattorro's Natural Reverberator Explained | PDF | Low Pass Filter, https://www.scribd.com/document/542133737/FinalTechnologyReview-DarioRizo

13. Reverberation processor based on absorbent all-pass filters, https://patents.google.com/patent/US7149314B2/en

14. Add reverberation to audio signal \- Simulink \- MathWorks, https://www.mathworks.com/help/audio/ref/reverberator.html

15. Reverb Topologists and Design \- Tom Erbe, https://tre.ucsd.edu/wp7/wp-content/uploads/2018/10/reverbtopo.pdf

16. Overview of Artificial Reverberation Techniques | PDF | Electrical, https://www.scribd.com/document/875602381/valimaki2012

17. Allpass Feedback Delay Networks \- arXiv, https://arxiv.org/pdf/2007.07337

18. Modal Decomposition of Feedback Delay Networks \- arXiv, https://arxiv.org/pdf/1901.08865

19. On Lossless Feedback Delay Networks \- arXiv, https://arxiv.org/pdf/1606.07729

20. Automatic Optimization of Feedback Delay Networks – SegfaultDSP, https://segfaultdsp.com/posts/reverb\_fdn/

21. Feedback Delay Networks: Echo Density and Mixing Time, https://www.researchgate.net/publication/311337136\_Feedback\_Delay\_Networks\_Echo\_Density\_and\_Mixing\_Time

22. Differentiable Feedback Delay Network for Colorless Reverberation, https://aaltodoc.aalto.fi/server/api/core/bitstreams/cf2d473f-b200-4e0d-8129-908a918247ec/content

23. Differentiable Feedback Delay Network For Colorless Reverberation, https://www.researchgate.net/publication/373756296\_Differentiable\_Feedback\_Delay\_Network\_For\_Colorless\_Reverberation

24. Circulant and elliptic feedback delay networks for artificial, https://www.researchgate.net/publication/3333489\_Circulant\_and\_elliptic\_feedback\_delay\_networks\_for\_artificial\_reverberation

25. Scattering in Feedback Delay Networks \- arXiv, https://arxiv.org/html/1912.08888v2

26. (PDF) Scattering in Feedback Delay Networks \- ResearchGate, https://www.researchgate.net/publication/342091767\_Scattering\_in\_Feedback\_Delay\_Networks

27. FDNTB: The Feedback Delay Network Toolbox, https://dafx2020.mdw.ac.at/proceedings/papers/DAFx2020\_paper\_53.pdf

28. Papers and code on delay-network-based artificial reverberation \[WIP\]., https://github.com/gdalsanto/delay-network-reverbs

29. The Stability of Multichannel Sound Systems With Time-Varying, https://www.researchgate.net/publication/305664096\_The\_Stability\_of\_Multichannel\_Sound\_Systems\_With\_Time-Varying\_Mixing\_Matrices

30. The Design of Matched Balanced Orthogonal Multiwavelets \- Frontiers, https://www.frontiersin.org/journals/applied-mathematics-and-statistics/articles/10.3389/fams.2021.785803/full

31. Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation, https://raw.githubusercontent.com/mlresearch/v235/main/assets/ma24a/ma24a.pdf

32. Orthonormal and unitary matrices \- Dan MacKinlay, https://danmackinlay.name/notebook/matrix\_orthonormal.html

33. Parametric Spring Reverberation Effect | Request PDF \- ResearchGate, https://www.researchgate.net/publication/230561075\_Parametric\_Spring\_Reverberation\_Effect

34. analytical approximations of fractional delays: lagrange, http://boytim.orgfree.com/asrc/icassp97\_paper.pdf

35. Allpass filter is not the same as fractional sample delay filter?, https://www.kvraudio.com/forum/viewtopic.php?t=569114

36. Simple design of fractional delay allpass filters \- ResearchGate, https://www.researchgate.net/publication/283101123\_Simple\_design\_of\_fractional\_delay\_allpass\_filters

37. 3.4 Design Methods for Fractional Delay Allpass Filters, http://users.spa.aalto.fi/vpv/publications/vesan\_vaitos/ch3\_pt3\_allpass.pdf

38. 4.5 Fractional Delay Operations with Allpass Filters, http://users.spa.aalto.fi/vpv/publications/vesan\_vaitos/ch4\_pt2\_allpass.pdf

39. Shimmer reverberation with nonlinear feedback delay networks, https://eprints.whiterose.ac.uk/id/document/4276219

40. DAFX \- Digital Audio Effects \- OLLINTEC, http://ollintec.com/procesamientodevoz/DAFX\_-\_Digital\_Audio\_Effects.pdf

41. DAFX \- Digital Audio Effects | Request PDF \- ResearchGate, https://www.researchgate.net/publication/241722015\_DAFX\_-\_Digital\_Audio\_Effects

42. KVR Audio Developer Challenge 2026 \- Free Plugins Competition, https://www.kvraudio.com/kvr-developer-challenge/2026/

43. Part III. Reference \- Csound, https://csound.com/docs/manual/PartReference.html

44. List of Accepted Papers, https://cmsworkshops.com/ICASSP2026/papers/accepted\_papers.php

45. Dafx : Digital Audio Effects \- SciSpace, https://scispace.com/pdf/dafx-digital-audio-effects-19nh43ug3t.pdf

46. TOMiC Drum Synth by Tomasz Głuc \- KVR Audio, https://www.kvraudio.com/product/tomic-drum-synth-by-tomasz-gluc

47. Grouped Feedback Delay Networks With Frequency-Dependent, https://acris.aalto.fi/ws/portalfiles/portal/113021875/Grouped\_Feedback\_Delay\_Networks\_With\_Frequency\_Dependent\_Coupling.pdf

48. Parameter Estimation of Digital Audio Effects from Spectrograms, https://cs231n.stanford.edu/2025/papers/text\_file\_840592643-CS231n\_Final\_Paper\_\_digital\_audio\_effects-4.pdf

49. Antiderivative Antialiasing Techniques in Nonlinear Wave Digital, https://re.public.polimi.it/retrieve/e0c31c12-807b-4599-e053-1705fe0aef77/21115.pdf

50. Comparative Study of Recurrent Neural Networks for Virtual Analog, https://arxiv.org/html/2405.04124v1

51. DAFx20in21 Programme \- Digital Audio Effects (DAFx), https://dafx2020.mdw.ac.at/DAFx20in21/index.html

52. Explicit Fixed-Point Computation of Nonlinear Delay-Free Loop, https://air.uniud.it/retrieve/e27ce0c9-73e4-055e-e053-6605fe0a7873/main\_oneColumn.pdf

53. Practical Considerations for Antiderivative Anti-Aliasing \- Medium, https://jatinchowdhury18.medium.com/practical-considerations-for-antiderivative-anti-aliasing-d5847167f510

54. SHIMMER REVERBERATION WITH NONLINEAR FEEDBACK, https://dafx.de/paper-archive/2026/papers/DAFx26\_paper\_05.pdf

55. Standard Functions \- Faust Libraries \- Grame, https://faustlibraries.grame.fr/standardFunctions/

56. dsp.DCBlocker \- Block DC component (offset) from input signal, https://www.mathworks.com/help/dsp/ref/dsp.dcblocker-system-object.html

57. (PDF) FAUST Domain Specific Audio DSP Language Compiled to, https://www.researchgate.net/publication/324642544\_FAUST\_Domain\_Specific\_Audio\_DSP\_Language\_Compiled\_to\_WebAssembly

58. News up to 2018 \- Faust Programming Language \- Grame, https://faust.grame.fr/community/news/

59. Optimizing the Code \- Faust Documentation \- ijc, https://ijc8.me/faustdoc/manual/optimizing/

60. DESIGNING SOFTWARE SYNTHESIZER PLUG-INS IN C++ with, https://dokumen.pub/designing-software-synthesizer-plug-ins-in-c-with-audio-dsp-2nbsped-9781000375053-1000375056.html

61. Audio Effects: Theory, Implementation And Application \[PDF\], https://vdoc.pub/documents/audio-effects-theory-implementation-and-application-3u9q01a7rpc0

62. Compiling Differentiable Audio Graphs to Real-Time DSP \- arXiv, https://arxiv.org/html/2606.21277v1