Generative Sound Engines and Continuous Timbral Morphing: DSP Architecture for Organic Spectral Evolution

Full report

On this page

Introduction

The pursuit of synthesizing sound that feels inherently organic, detailed, and alive stands as one of the preeminent challenges in digital signal processing (DSP) and generative audio design. Traditional synthesizer architectures have historically relied on static oscillator presets, deterministic low-frequency oscillators (LFOs), and fixed filter topologies. While these structures are highly effective for rhythmic, fast-attack, and melodic contexts where event density masks timbral stasis, they reveal their artificiality when tasked with sustaining single notes or continuous textures over extended timeframes. Over periods spanning seconds, minutes, or hour-long generative sessions, static waveforms induce auditory fatigue, and strictly periodic modulation betrays the underlying digital determinism of the system.

To create a continuously morphing timbre that mimics the infinite, chaotic complexity of physical acoustic phenomena, a generative sound engine must transcend linear interpolation between wavetables. It must operate across multiple intersecting domains: additive partial manipulation, stable non-linear filter evolution, chaotic modulation systems, and perceptually guided feature interpolation. The ultimate goal is to produce spectral evolution that maintains continuous auditory interest without relying on constant event density or overt rhythmic variation.

This report provides an exhaustive analysis of the DSP structures, psychoacoustic principles, and algorithmic strategies required to engineer a generative sound engine capable of organic spectral evolution. It investigates the mechanics of harmonic versus inharmonic spectra, advanced filter-bank morphing, zero-delay feedback topologies, and anti-aliasing techniques. Furthermore, it establishes a framework for abstracting these complex mathematical transformations into intuitive, artistic macro-controls via a high-level AudioCharacter architecture, determining which parameters should remain independent controls and which should function as internal, deterministic dimensions, while maintaining strict DSP stability and backward compatibility.

The Psychoacoustics of Slow Spectral Evolution

To maintain auditory interest in long-form generative sessions, the sound engine must exploit the mechanisms of human auditory perception, specifically addressing the neurological phenomena of "repetition suppression" and "auditory change deafness." Understanding these mechanisms is critical for designing spectral transformations that feel alive rather than fatiguing.

Repetition Suppression and 1/f Fractal Dynamics

Repetition suppression describes the phenomenon whereby the central nervous system exhibits a sparser, attenuated neuronal response to repeated or static stimuli1. Adaptation in the auditory cortex actively reduces neuronal responses to unchanging sounds, producing a highly efficient representation that filters out predictable background noise to prioritize novel environmental cues3. Consequently, a static oscillator preset, regardless of how harmonically rich it is at its onset, will rapidly be perceived as dull, synthetic, or "lifeless" as the brain adapts to its invariant statistics.

Conversely, biological systems and natural physical phenomena continuously exhibit self-similar, scale-invariant micro-fluctuations, frequently characterized by [Figure omitted from source export] ("pink noise") dynamics4. In motor control, neurological activity, and acoustic environments, fractal complexity associated with healthy, adaptive systems takes on a pink noise character, balancing complete randomness (white noise) and excessive predictability6. By injecting [Figure omitted from source export] fractal noise into the modulation matrices of a sound engine, the spectral evolution directly mirrors natural phenomena4. Because these fluctuations contain energy across all octaves but heavily favor slow, drifting changes, the auditory system is constantly fed subtle novel micro-information. This prevents complete repetition suppression while avoiding the jarring, unmusical discontinuity of white noise modulation.

Auditory Change Deafness and Sub-Threshold Morphing

Auditory change deafness refers to the well-documented inability of human listeners to detect substantial changes in complex auditory scenes when their attention is not explicitly directed to the changing element, or when the change occurs gradually over an extended period7. In generative soundscapes with low event density, this perceptual limitation can be leveraged artistically to create expansive spectral evolutions.

By executing macroscopic spectral transformations—such as shifting the spectral centroid, morphing formants, or altering the spectral spread—over very long timeframes (ranging from tens of seconds to several hours), the engine alters the fundamental emotional character of the sound without triggering the listener's transient-detection mechanisms7. The sound feels "alive" because it is perpetually in a state of becoming. The transformations remain sub-perceptual in the short term, allowing the listener to inhabit the texture without being distracted by abrupt shifts, thus bypassing auditory fatigue entirely.

Spectral Structures and Morphing Modalities

To achieve continuous, organic timbral morphing, the generative engine must manipulate the spectrum at granular, partial, and macroscopic levels. Direct cross-fading of two static audio signals does not constitute morphing; it results in a perceptual superposition—two distinct sounds playing simultaneously at varying amplitudes. True morphing requires the interpolation of parameterized spectral features derived from rigorous analysis8.

Harmonic versus Inharmonic Spectra and Harmonicity

The physical foundation of timbre lies in the frequency relationship between an oscillator's partials. Harmonic spectra, where partials are exact integer multiples of the fundamental frequency ([Figure omitted from source export]), produce the clear, unambiguous sense of pitch characteristic of ideal strings and vocal tracts. Inharmonic spectra, where partials deviate from integer multiples, produce the complex, metallic, bell-like, or percussive timbres characteristic of struck metals, stiff rods, and asymmetrical bells10.

A generative engine must treat "Harmonicity" not as a binary state, but as a fluid, modulatable dimension. By applying a dispersion coefficient to the partials—mimicking the physical stiffness of a resonating string—the engine can smoothly warp a harmonic spectrum into an inharmonic one11. The wave equation for stiff strings introduces a complex dispersive partial differential equation, where the frequency of the [Figure omitted from source export]\-th partial is modeled as:

[Figure omitted from source export]

In this equation, [Figure omitted from source export] serves as the inharmonicity factor11. Modulating [Figure omitted from source export] continuously over time transforms a warm, organic, harmonic string texture into a dissonant, metallic chime, providing an immensely powerful tool for slow spectral evolution.

Partial Drift and Additive Activation/Deactivation

In acoustic instruments, partials do not remain perfectly phase-locked, nor are they statically amplitude-scaled. They exhibit independent micro-fluctuations in both frequency and amplitude, commonly referred to as partial drift14. An advanced additive synthesis engine models this organic instability by assigning independent [Figure omitted from source export] drift generators to the amplitude and pitch of every individual partial, breaking the rigid phase relationships inherent in digital oscillators.

To manage the CPU implications of calculating hundreds of independent, drifting sine waves, Polynomial Additive Sound Synthesis (PASS) can be employed. PASS approximates sine functions with computationally cheap polynomials and updates coefficients only when necessary via a priority queue, operating with a complexity proportional to the sum of the frequencies rather than strictly the number of oscillators15.

Furthermore, structured pruning must be implemented for additive partial activation and deactivation. Sub-threshold partials—those completely masked by louder adjacent frequencies according to psychoacoustic masking curves, or those falling below the absolute threshold of hearing—must be dynamically deactivated to reclaim processing cycles15. The system continuously evaluates the masking threshold of the total spectral output using an Equivalent Rectangular Bandwidth (ERB) filter-bank model, formulated as [Figure omitted from source export]18. If a morphing partial drifts below the psychoacoustic masking threshold, its amplitude envelope gracefully fades to zero, and the oscillator is culled from the computation loop, allowing the engine to render incredibly dense textures efficiently.

High-Level Spectral Descriptors

To guide the morphing process organically, the engine relies on high-level audio descriptors that map directly to human perception. Morphing between two timbres is achieved by interpolating these descriptor values and continuously updating the underlying DSP parameters to match the target features9.

• Spectral Centroid: The amplitude-weighted mean frequency of the spectrum. It correlates strongly with the perceptual dimension of "Brightness" or "Sharpness."
• Spectral Spread: The variance or bandwidth of energy around the centroid. A low spread indicates a pure, sine-like tone, while a high spread indicates a rich, broadband texture.
• Spectral Flatness: Also known as Wiener entropy, this measures the ratio of the geometric mean to the arithmetic mean of the power spectrum. It distinguishes between tonal (harmonic, low flatness) and noise-like (high flatness) signals.
• Spectral Flux: The rate of change of the magnitude spectrum from one frame to the next. In a generative context, targeting a specific spectral flux allows the engine to enforce a consistent level of "Motion" or internal activity, even during extremely slow macroscopic shifts.

Formants and LSF Morphing

Formants represent the resonant peaks of an acoustic system's transfer function (its filter) and define the vowel-like characteristics and the physical "body" of a sound. When morphing between two timbres (e.g., transitioning from a dark, muffled drone to a bright, vocalic texture), simply crossfading the magnitudes of a Short-Time Fourier Transform (STFT) causes the original formants to fade out in place while new ones fade in. This is perceptually invalid and destroys the illusion of a singular, transforming acoustic body9.

Instead, spectral envelopes must be extracted and their temporal and spectral events aligned using Dynamic Time Warping (DTW) or Dynamic Frequency Warping (DFW) to accurately match and shift spectral peaks across the frequency axis20. However, interpolating standard Linear Predictive Coding (LPC) coefficients to achieve formant shifting frequently results in highly unstable filters that blow up into harsh digital noise20.

To guarantee filter stability during the interpolation of these resonant peaks, LPC coefficients must be transformed into Line Spectral Frequencies (LSF), also known as Line Spectral Pairs (LSP)20. LSFs represent the roots of two auxiliary polynomials, [Figure omitted from source export] (symmetric) and [Figure omitted from source export] (antisymmetric), derived from the inverse filter. A critical mathematical property of LSFs is that these roots lie strictly on the unit circle and are perfectly interlaced24. When interpolating between the LSFs of two distinct timbres, as long as the roots remain on the unit circle and their interlacing property is preserved, the resulting intermediate synthesis filter is mathematically guaranteed to be minimum-phase and unconditionally stable25. This allows for the smooth, continuous morphing of formants, driving profound changes in the spectral centroid and overall timbre without generating unstable resonant feedback9.

Filter-Bank Morphing, Moving Notches, and Comb Structures

Beyond macroscopic formants, generating an organic texture requires simulating the highly complex, microscopic resonant behaviors of physical bodies, such as wooden soundboards or metal plates, which exhibit modal resonances—densely packed, high-Q resonant peaks10.

A highly effective method for continuous morphing is to route a harmonically rich, drifting excitation signal through a dynamically morphing filter-bank. By introducing moving notches (via time-varying all-pass filters or cascading phaser topologies) and comb structures (via short, modulated feedback delay lines), the engine creates sweeping spectral voids. When the frequencies of these notches are modulated by a chaotic system, they simulate the continuously shifting phase relationships of a sound source moving through a complex physical space, creating the illusion of three-dimensionality28.

Modal Resonances, Convolution Textures, and Resonator Excitation

While non-diffuse resonators (like a struck marimba) can be modeled using parallel banks of second-order sections, diffuse resonance (such as the body of a cello, a heavily damped metallic plate, or a reverberant room) requires dense feedback delay networks (FDNs) or multi-dimensional digital waveguide meshes (DWM)27.

To produce organic timbres, the excitation of these resonators is paramount. Rather than relying on simple impulse trains, the generative engine utilizes continuous, colored noise excitation27. By convolving a dry additive signal with time-varying impulse responses (IRs) of physical spaces or materials, the engine imparts profound physical realism to the sound. Furthermore, by dynamically interpolating between different IRs in the frequency domain, the "Texture" of the sound can smoothly morph from the resonance of a wooden enclosure to the flutter-echo of a metallic pipe33.

Synthesis Paradigms for Organic Evolution

To translate spectral parameters and descriptors into an active, audible soundscape, the generative engine must employ specific synthesis algorithms that allow for extreme, continuous modification without introducing digital artifacts, such as phase discontinuities, clicks, or aliasing.

Wavetable Interpolation and Phase Modulation

Wavetable synthesis provides an efficient method for generating harmonically rich spectra by reading single-cycle waveforms from memory. Smooth timbral transitions require multi-dimensional crossfading between adjacent waveforms in a 3D wavetable matrix. However, if the waveforms in the matrix are not perfectly phase-aligned, linear interpolation causes destructive phase cancellation, resulting in a sudden, unnatural drop in spectral energy and artificial amplitude dipping34.

To resolve this, all wavetables must be pre-processed mathematically to be minimum-phase. Minimum-phase signals have all their energy concentrated as close to the beginning of the impulse as possible, ensuring that phase relationships are aligned across the entire interpolation axis, allowing for seamless spectral morphing without volume dips26.

Phase Modulation (PM) and Frequency Modulation (FM) provide immense spectral variation with minimal CPU overhead. By mapping a slow, chaotic modulator to the FM index (the modulation depth), the Bessel-function-derived sidebands expand and contract continuously36. This simulates the non-linear energy transfer of a physical instrument being bowed or blown with varying, breath-like pressure. Utilizing loopback FM (where an oscillator modulates its own frequency via a time-varying delay line) introduces complex, evolving noise-like components into the spectrum, highly dependent on the fractional delay interpolation used36.

Waveshaping, Anti-Aliasing (ADAA), and Noise Coloration

Waveshaping injects harmonics by passing a signal through a non-linear transfer function, such as Chebyshev polynomials or hyperbolic tangent functions ([Figure omitted from source export]). This is essential for generating warmth, drive, and edge. However, dynamic waveshaping of high-frequency components mathematically guarantees the generation of harmonics above the Nyquist frequency ([Figure omitted from source export]). In a digital system, these frequencies reflect back into the audio band as inharmonic aliasing distortion, which sounds harsh, digital, and completely destroys the illusion of an organic timbre37.

To maintain pristine audio quality over long, evolving sessions, traditional oversampling (inserting zeros and low-pass filtering) is often insufficient or introduces unacceptable CPU overhead and phase-cancellation issues due to steep brick-wall filters26. Instead, Antiderivative Anti-Aliasing (ADAA) is required40.

ADAA analytically approximates continuous-time convolution by evaluating the antiderivative of the non-linear waveshaping function. For a continuous non-linear function [Figure omitted from source export] with an exact analytical antiderivative [Figure omitted from source export], the first-order ADAA output for an input sample [Figure omitted from source export] is computed as:

[Figure omitted from source export]

When the denominator [Figure omitted from source export] approaches zero (an ill-conditioned state), the equation falls back to the original function [Figure omitted from source export] or uses a Taylor series expansion to avoid division by zero43. This process heavily suppresses the first set of aliased harmonics, producing a clean, artifact-free morphing distortion. Utilizing second-order ADAA (which requires the second antiderivative) in conjunction with moderate [Figure omitted from source export] oversampling can achieve over [Figure omitted from source export] of alias suppression with minimal CPU impact, preserving the organic purity of the sound41.

Noise coloration further enhances this organic nature. Injecting [Figure omitted from source export] pink noise directly into the signal path creates a textural bed that mimics physical wind, breath, or the friction of a bow on a string45. The Voss-McCartney algorithm efficiently generates true pink noise by summing multiple white-noise generators sampled at successively halved rates, producing the characteristic [Figure omitted from source export] spectral tilt that the human ear perceives as a balanced, natural rushing sound45.

Granular Timbre and Time-Stretching

Granular synthesis deconstructs audio into microscopic grains (typically 10 to 100 milliseconds) and redistributes them in time and space. By continuously modulating the grain size, grain density, and playback position with chaotic attractors, the engine can stretch microscopic acoustic textures into infinite, evolving drones.

Strict windowing functions (e.g., Gaussian, Hann, or Tukey windows) must be applied to the amplitude envelope of every single grain to prevent phase discontinuities and clipping artifacts at the grain boundaries49. Spectral granular synthesis advances this further by extracting only the magnitude spectrum of the grains in the frequency domain and reconstructing the phase entirely via spectrogram inversion techniques. This allows for extreme time-stretching and pitch-shifting without the metallic, robotic artifacts commonly associated with phase-vocoder smearing, resulting in a pristine, atmospheric timbre51.

High-Level AudioCharacter Macro-Dimensions

A core requirement of a modern generative engine is exposing immensely complex DSP ecosystems to users through artistic, intuitive, and transparent parameters. Users should not be required to manually calculate LSF coefficients, structure ADAA derivatives, or tune [Figure omitted from source export] fractal exponents. Instead, the engine relies on an AudioCharacter framework—a multidimensional abstraction layer.

This strategy divides control into "Independent Controls" (macros exposed to the user) and "Internal Dimensions" (calculated deterministically under the hood based on the complex interaction of the independent controls).

Mapping the Concept to DSP Architecture

 

Extensibility and Backward Compatibility Strategy

To ensure the AudioCharacter framework remains fully backward-compatible as the generative engine evolves over years of updates, it must be designed using deterministic reference metrics and strict versioned parameter bounding.

If a future iteration of the engine introduces a new filter topology or an updated granular windowing function, the AudioCharacter framework continues to rely on unitless normalized floating-point ranges ([Figure omitted from source export]) for all saved state data. A versioned abstraction layer maps these normalized values to physical limits (e.g., [Figure omitted from source export] to [Figure omitted from source export]) based on the specific version of the DSP algorithm initialized at load time. This architecture guarantees that an AudioCharacter preset saved in version 1.0 translates mathematically equivalently to the updated engine in version 2.0, preserving the artist's original timbral intent without breaking past generative sessions.

DSP Robustness, Artifacts, and CPU Optimization

Sustaining a continuously evolving, harmonically dense sound over long generative sessions introduces severe numerical stability challenges. The DSP architecture must be mathematically fortified against edge cases, unbounded state accumulation, and computational bottlenecks.

Zero-Delay Feedback (ZDF) SVF and Extreme Q

Traditional digital IIR filters derived via the standard bilinear transform suffer from catastrophic delay-free loop problems when resonance (feedback) is introduced. The inherent one-sample delay in the digital feedback path causes severe phase errors at high frequencies, high-frequency warping (where the filter fails to track accurately near the Nyquist limit), and explosive instability under fast modulation54. For continuous, organic modulation where filter parameters are swept at audio rates, the engine must employ Topology-Preserving Transform (TPT) or Zero-Delay Feedback (ZDF) filters, a methodology extensively formalized by Vadim Zavalishin in The Art of VA Filter Design54.

By substituting analog integrators with digital trapezoidal integrators and algebraically solving the implicit feedback loop within the same sample period, the current output is calculated without relying on delayed feedback states54. The core calculation for the high-pass node in a ZDF SVF becomes:

[Figure omitted from source export]

where [Figure omitted from source export] (providing exact frequency pre-warping), [Figure omitted from source export] is the damping factor ([Figure omitted from source export]), and [Figure omitted from source export] are the memory states of the integrators54. This ZDF SVF architecture guarantees absolute mathematical stability even when the cutoff frequency ([Figure omitted from source export]) is modulated at extreme rates, and it provides stable, analog-matched self-oscillation at maximum [Figure omitted from source export] settings, which is essential for sweeping resonant peaks54. To introduce organic warmth into this structure, fast Padé approximants of the [Figure omitted from source export] function can be placed inside the integrator loops, requiring iterative Newton-Raphson solving or linearized feedback saturation to prevent runaway54.

Aliasing, Denormals, Ultrasonic Content, and Clipping

• Aliasing: Managed primarily through the aforementioned Antiderivative Anti-Aliasing (ADAA) algorithms. Where ADAA is mathematically insufficient (such as in extremely hard clipping or complex FM feedback), [Figure omitted from source export] or [Figure omitted from source export] oversampling using linear-phase FIR interpolation filters is dynamically engaged to push fold-over reflections out of the audible spectrum26.
• Denormals (Subnormal Numbers): As partials decay or resonant feedback delay networks ring out over long ambient sessions, IEEE 754 floating-point numbers can drop to subnormal levels (e.g., [Figure omitted from source export]). When CPUs encounter denormals, they trap to microcode to calculate the extremely small values, resulting in performance penalties of up to 100x and causing catastrophic audio dropouts (CPU spikes)57. The sound engine must strictly enforce hardware-level Flush-to-Zero (FTZ) and Denormals-Are-Zero (DAZ) flags at the thread level, guaranteeing deterministic CPU usage regardless of the signal's decay state57.
• Ultrasonic Content & Clipping: Chaotic attractors (like the Lorenz system) can inadvertently drive SVF cutoffs into the ultrasonic range, or cause unpredictable constructive phase alignment across hundreds of partials that exceeds [Figure omitted from source export]. The engine utilizes a rigorously bounded domain for all parameter mappings and employs a soft-clipping saturation stage equipped with ADAA at the master bus to gracefully absorb transient spikes without harsh digital clipping54.

CPU Implications, Vectorization, and Pruning

Long-duration generative music implies dense polyphonic layering and highly complex spectral calculation. To optimize CPU consumption, the architecture must abandon naive [Figure omitted from source export] static additive synthesis15. Instead, it relies on structured pruning and Single Instruction, Multiple Data (SIMD) vectorization17.

The system continuously evaluates the ERB masking threshold of the total spectral output. If a morphing partial drifts below the psychoacoustic masking threshold of adjacent, dominant partials, it is gracefully faded out, and its underlying oscillator math is entirely culled from the computation loop15. By combining structured pruning, PASS polynomial approximations, and DAZ/FTZ flags, the engine sustains massive spectral complexity while operating comfortably within real-time computational limits.

Deterministic Evaluation Metrics and Perceptual Frameworks

To ensure the deterministic quality and perceptual validity of the generative engine during development and tuning, rigorous subjective and objective evaluation metrics must be utilized.

Objective Metrics: The engine evaluates morphing linearity, spectral integrity, and artifact suppression using standardized algorithms such as the Perceptual Evaluation of Audio Quality (PEAQ, ITU-R BS.1387) and the Hearing-Aid Audio Quality Index (HAAQI)59. These tools measure structural fidelity, temporal fine structure preservation, and spectral envelope integrity against a baseline60. Morphing linearity is mathematically quantified by observing the trajectory of the spectral centroid and LSF vectors; a perceptually ideal morph will present a strictly linear trajectory in the ERB-warped frequency domain, ensuring smooth transitions without sudden drop-offs9.

Subjective Metrics: MUSHRA (Multiple Stimuli with Hidden Reference and Anchor) testing protocols are employed to validate the "organic" feel of the morphing algorithms59. Listeners rate the continuous morph against a hidden reference (e.g., the unmodulated, static high-resolution waveform) and an anchor (a highly aliased, low-pass filtered version). This double-blind methodology ensures that the ADAA algorithms, ZDF filters, and 1/f modulation matrices successfully suppress digital artifacts and emulate analog organicism to a degree indistinguishable by the human ear.

Conclusion

The architecture detailed in this report establishes a comprehensive, mathematically rigorous DSP framework for a generative sound engine capable of organic, continuously evolving timbres. By abandoning static waveforms and deterministic LFOs, and instead embracing [Figure omitted from source export] noise coloration, chaotic Lorenz attractor modulation, and Polynomial Additive Sound Synthesis with dynamic psychoacoustic pruning, the engine completely evades the neurological pitfalls of auditory repetition suppression.

The strict integration of Line Spectral Frequency interpolation, Zero-Delay Feedback state variable filters, and Antiderivative Anti-Aliasing ensures that the engine remains mathematically stable, minimum-phase, and artifact-free, even under extreme resonance, continuous formant morphing, and complex non-linear waveshaping. Finally, by abstracting these deep mathematical operations into the perceptually grounded, version-controlled AudioCharacter macro-dimensions, the engine bridges the divide between mathematically demanding DSP and intuitive, transparent sound design, providing a scalable, efficient platform for infinite acoustic generation.

Works cited

1. Music, repetition and time perception \- uniPUB, https://unipub.uni-graz.at/obvugrhs/download/pdf/2845597

2. Neuroscience and New Music: Assessing Behavioral and Cerebral, https://arodes.hes-so.ch/record/11144/files/James\_2022\_neuroscience\_book.pdf

3. Adaptation in auditory processing \- PMC \- NIH, https://pmc.ncbi.nlm.nih.gov/articles/PMC9829473/

4. Colors of noise \- Grokipedia, https://grokipedia.com/page/Colors\_of\_noise

5. The Readiness-Efficiency Coupling: Synergy of High Preparatory, https://www.tandfonline.com/doi/full/10.1080/00222895.2026.2670520

6. Synergy of High Preparatory and Reduced Execution Fractal Complex, https://research.utwente.nl/files/541077343/The\_Readiness-Efficiency\_Coupling\_Synergy\_of\_High\_Preparatory\_and\_Reduced\_Execution\_Fractal\_Complexity\_in\_Motor\_Learning.pdf

7. Directed Attention Eliminates 'Change Deafness' in Complex, https://www.researchgate.net/publication/7779643\_Directed\_Attention\_Eliminates\_'Change\_Deafness'\_in\_Complex\_Auditory\_Scenes

8. Timbre Morphing using the Modal Distribution, https://www.dafx.de/paper-archive/2002/DAFX02\_Lysaght\_Timoney\_timbre\_morphing.pdf

9. Sound Morphing by Audio Descriptors and parameter interpolation, https://www.mcgill.ca/mpcl/files/mpcl/kazazis\_2016\_dafx.pdf

10. Efficient Modeling and Synthesis of Bell-like Sounds, https://www.dafx.de/papers/DAFX02\_Karjalainen\_Valimaki\_Esquef\_bell-like\_sounds.pdf

11. Hammer-String Simulation with Time-Variant Parameters, https://www.dafx.de/paper-archive/2023/DAFx23\_paper\_18.pdf

12. Physical Modelling of the Bowed String and Applications to Sound, https://www.acoustics.ed.ac.uk/wp-content/uploads/Theses/Desvages\_Charlotte\_\_PhDThesis\_UniversityOfEdinburgh\_2018.pdf

13. (PDF) The wave equation for stiff strings and piano tuning, https://www.researchgate.net/publication/301838415\_The\_wave\_equation\_for\_stiff\_strings\_and\_piano\_tuning

14. Additive Synthesis \- The Csound FLOSS Manual, https://flossmanual.csound.com/sound-synthesis/additive-synthesis

15. Fast Additive Sound Synthesis Using Polynomials, https://www.dafx.de/paper-archive/2006/papers/p\_181.pdf

16. Fast Additive Sound Synthesis Using Polynomials, https://www.dafx.de/paper-archive/details/Pf-LRikIFjeRRJHAdGLB-Q

17. Inference-Time Structured Pruning for Real-Time Neural ... \- DAFx25, https://dafx25.dii.univpm.it/wp-content/uploads/2025/09/DAFx25\_paper\_78.pdf

18. Modulation vocoder for analysis, processing and synthesis of audio, https://edocs.tib.eu/files/e01dh11/663198445.pdf

19. MULTIPLE-FEATURE FUSION BASED ONSET DETECTION FOR, https://ismir2008.ismir.net/papers/ISMIR2008\_127.pdf

20. Musical Instrument Sound Morphing Guided by Perceptually, http://articles.ircam.fr/textes/Caetano13a/index.pdf

21. Psychophysical scaling of timbre-related audio descriptors, https://www.mcgill.ca/mpcl/files/mpcl/kazazis\_2020\_phdthesis.pdf

22. An Educational Guide through the FMP Notebooks for Teaching and, https://www.mdpi.com/2624-6120/2/2/18

23. Improving Speech Intelligibility through Spectral Style Conversion, https://digitalcollections.ohsu.edu/record/9298/files/Dinh.Tuan.2021.pdf

24. The Computation of Line Spectral Frequencies Using Chebyshev, http://hil.t.u-tokyo.ac.jp/\~kameoka/SAP/papers/Kabal1986\_\_The\_Computation\_of\_Line\_Spectral\_Frequencies\_Using\_Chebyshev\_Polynomials.pdf

25. A one-dimensional search method with stable 1-norm solution for, https://pubs.aip.org/asa/jasa/article/142/2/EL170/640348/A-one-dimensional-search-method-with-stable-1-norm

26. (PDF) Antiderivative Antialiasing for Arbitrary Waveform Generation, https://www.researchgate.net/publication/362628103\_Antiderivative\_Antialiasing\_for\_Arbitrary\_Waveform\_Generation

27. perceptual resonators for interactive worlds \- ICAD, https://icad.org/Proceedings/2002/Menzies2002b.pdf

28. DAFX \- Digital Audio Effects | Request PDF \- ResearchGate, https://www.researchgate.net/publication/242019278\_DAFX\_-\_Digital\_Audio\_Effects

29. Differentiable grey-box modelling of phaser effects using frame, https://www.dafx.de/paper-archive/2023/DAFx23\_paper\_38.pdf

30. Acoustic modeling using the digital waveguide mesh, https://eprints.whiterose.ac.uk/id/eprint/3708/1/murphyd2.pdf

31. Physical Modelling of the Vocal Tract with the 2D Digital Waveguide, https://www-users.york.ac.uk/\~dtm3/Download/JackThesis.pdf

32. (PDF) Perceptual Evaluation of Rolling Sound Synthesis, https://www.researchgate.net/publication/233512880\_Perceptual\_Evaluation\_of\_Rolling\_Sound\_Synthesis

33. Scalable and efficient room acoustic modeling using delay networks, https://openresearch.surrey.ac.uk/view/pdfCoverPage?instCode=44SUR\_INST\&filePid=13223435420002346\&download=true

34. Music Production Behaviour Modelling, https://qmro.qmul.ac.uk/xmlui/bitstream/handle/123456789/89253/PhD\_Dissertation\_\_\_Corrections.pdf

35. PAPERS \- Analysis and Synthesis of Tones by Spectral Interpolation, https://www.cs.cmu.edu/\~./rbd/papers/Spectral-JAES-1990.pdf

36. Program \- Digital Audio Effects Conference DAFx26 \- MIT, https://dafx26.mit.edu/program/

37. Reducing the aliasing of nonlinear waveshaping using continuous, https://dafx.de/paper-archive/2016/dafxpapers/20-DAFx-16\_paper\_41-PN.pdf

38. Oversampling for Nonlinear Waveshaping: Choosing the Right Filters, https://scispace.com/pdf/oversampling-for-nonlinear-waveshaping-choosing-the-right-1qy7m7qoeu.pdf

39. Introduction to Oversampling for Alias Reduction \- Nick Thompson, https://www.nickwritesablog.com/introduction-to-oversampling-for-alias-reduction/

40. antialiased \- Faust Libraries \- Grame, https://faustlibraries.grame.fr/libs/aanl/

41. KERN WARM vs puffer:fish: two approaches to saturation, https://kernaudio.io/guides/saturation/warm-vs-pufferfish

42. Minimal-Control DSP for Mix-Bus Processing, https://projekter.aau.dk/projekter/files/795734201/Karl\_EmilHald\_Master\_Thesis\_2025\_2.pdf

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

44. 19 International Conference on Digital Audio Effects \- DAFx, https://dafx16.vutbr.cz/files/DAFx16\_proceedings.pdf

45. NOISE2, a white and pink noise source in a single chip \- Electric Druid, https://electricdruid.net/noise2-white-pink-noise-source/

46. Pink noise generator \- JUCE Forum, https://forum.juce.com/t/pink-noise-generator/40013

47. PNoiseGenerator — Acoular 26.01 documentation, https://acoular.org/acoular/v26.01/api\_ref/generated/generated/acoular.signals.PNoiseGenerator.html

48. Ear Training Frequency Tool — Learn to Identify EQ, https://frequencydetector.com/ear-training-frequency-tool/

49. The Basics of Granular Synthesis \- iZotope, https://www.izotope.com/community/blog/the-basics-of-granular-synthesis

50. Granular Synthesis Explained \- ACE Studio, https://acestudio.ai/blog/what-is-granular-synthesis/

51. (PDF) Spectral Granular Synthesis \- ResearchGate, https://www.researchgate.net/publication/326316857\_Spectral\_Granular\_Synthesis

52. VORTESSA nonlinear generative system for unstable percussion, https://www.elektronauts.com/t/vortessa-nonlinear-generative-system-for-unstable-percussion-feedback-dissipative-synthesis/249875

53. Creating Organic Modulations in Ableton with AZ Lorenz XYZA 1.0, https://lame.buanzo.org/max4live\_blog/unleashing-the-chaos-creating-organic-modulations-in-ableton-with-az-lorenz-xyza-10.html

54. Zero-Delay Feedback Filters: Stable Digital Emulations of Analog, https://synthengineer.com/blog/zero-delay-feedback-filters

55. Vadim Zavalishin \- A different kind of digital filter theory \- YouTube, https://www.youtube.com/watch?v=zPzCLqkQnr0

56. An 4VirtualAnalogFilters | PDF \- Scribd, https://www.scribd.com/document/607788865/An-4VirtualAnalogFilters

57. Denormals and Ardour \- Blog, https://discourse.ardour.org/t/denormals-and-ardour/78129

58. Sarmays Super Saturator — Band-Focused Saturation Plugin, https://sarmays.com/products/SuperSaturator

59. Neural Audio Effect Modelling Strategies for a Spring Reverb \- Zenodo, https://zenodo.org/records/8380480/files/Francesco-Papaleo-Master-Thesis-2023.pdf?download=1

60. The Hearing-Aid Audio Quality Index (HAAQI) \- ResearchGate, https://www.researchgate.net/publication/286479699\_The\_hearing-aid\_audio\_quality\_index\_HAAQI

61. AudioLabs \- Publications \- International Audio Laboratories Erlangen, https://www.audiolabs-erlangen.de/fau/professor/herre/Publications

62. Detailed Program \- DAFx25, https://dafx25.dii.univpm.it/detailed-program/

63. Real-Time and High-Quality Musical Audio Streaming Over IP Luca, https://iris.unitn.it/retrieve/be8b4498-a953-484f-b36a-7b0913c66f17/PhD%20Thesis%20Luca%20Vignati.pdf