Audio systems, ranging from analog effects and guitar amplifiers to modern MEMS loudspeakers, are inherently multiphysics and nonlinear devices. The nonlinear interactions among different physical domains, together with phenomena such as magnetic or piezoelectric hysteresis, are key to the distinctive behavior and sonic character of these systems. Capturing these effects in a lightweight digital representation is essential for both Virtual Analog (VA) modeling, which aims to faithfully emulate analog audio circuits in the digital domain, and for the design and optimization of audio transducers, where accurate physical models support the development of Digital Signal Processing (DSP) algorithms for performance enhancement. Lumped-Element Models (LEMs) provide a compact and physically interpretable framework for describing such systems by exploiting analogies between physical domains and representing them as equivalent electrical networks of interconnected circuit elements. These models are particularly well suited to audio applications, as they can be implemented efficiently in the discrete-time domain and interfaced with DSP structures. However, LEMs have limited expressive power when it comes to capturing strongly nonlinear or hysteretic effects, and their analytical treatment often leads to implicit equations that are difficult to solve in real-time. To address these challenges, this dissertation extends the Wave Digital Filter (WDF) formalism into a unified, differentiable modeling and implementation framework. In this approach, circuit elements and connection networks are described as input-output blocks characterized by scattering relations, where port voltages and currents are expressed as linear combinations of incident and reflected waves through the introduction of a reference port resistance. The choice of this free parameter per port provides important numerical advantages, ensuring stability, modularity, and computational efficiency. Building on this modular structure, in this thesis we introduce neural network-based Wave Digital (WD) blocks for representing the nonlinear scattering equations of one-port and multi-port elements. Both explicit and implicit hybrid WD structures are investigated: explicit realizations enable real-time simulation of circuits containing a single nonlinear neural element, while implicit realizations address more complex circuits with multiple nonlinear WD blocks using iterative solvers. Both standard and hybrid WD structures can be simulated in the discrete-time domain within an Automatic Differentiation (AD)-based framework, which enables the computation of exact derivatives of the simulated response with respect to the model parameters. This allows the gradient-based optimization of both circuit parameters and neural parametric models directly from experimental measurements. In this setting, the parameters of linear and nonlinear LEMs can be refined according to a supervised learning paradigm, with neural WD blocks trained within the same differentiable simulation loop. The proposed modeling framework is applied to two main areas. In VA modeling, hybrid WDF structures are used to emulate nonlinear audio circuits achieving high accuracy and real-time efficiency; in piezoelectric MEMS loudspeaker modeling, differentiable LEMs combining WDF implementations and neural network-based processing blocks are used to capture the nonlinear electro-mechano-acoustic behavior of these transducers, supported by data-driven parameter identification directly from acoustic measurements. Finally, building upon the established LEM modeling framework, we address circuit inversion and transducer virtualization, introducing a systematic method for deriving inverse circuit representations based on nullors. This enables the design of model-based DSP algorithms that digitally compensate or modify the behavior of audio transducers through cascaded direct and inverse systems. The approach is applied to a piezoelectric MEMS loudspeaker equalization case study, demonstrating accurate response flattening while preserving the interpretability of the underlying physical model.
I sistemi audio, che spaziano dagli effetti analogici e dagli amplificatori per chitarra fino ai più moderni altoparlanti MEMS, sono dispositivi intrinsecamente nonlineari e multifisici. Le interazioni nonlineari tra diversi domini fisici, insieme a fenomeni come l'isteresi magnetica o piezoelettrica, svolgono un ruolo chiave nel determinarne il comportamento fisico e la loro peculiare impronta sonora. Catturare tali effetti tramite modelli digitali efficienti risulta essenziale sia nell'ambito del Virtual Analog (VA) modeling, che mira ad emulare fedelmente i circuiti audio analogici nel dominio digitale, sia nella progettazione ed ottimizzazione di trasduttori elettroacustici, dove modelli fisici accurati supportano lo sviluppo di algoritmi di Digital Signal Processing (DSP) finalizzati al miglioramento delle prestazioni. I Lumped-Element Models (LEMs) forniscono strutture compatte e fisicamente interpretabili per descrivere tali sistemi, sfruttando analogie elettriche tra domini fisici differenti e rappresentandoli come reti equivalenti di elementi circuitali interconnessi. Questi modelli risultano particolarmente adatti alle applicazioni audio, poiché possono essere implementati in maniera efficiente nel dominio tempo-discreto ed integrati con altre strutture DSP. Tuttavia, i LEM presentano limitazioni nella capacità di rappresentare effetti fortemente nonlineari o isteretici, il cui trattamento analitico può condurre alla definizione di equazioni implicite difficili da risolvere in tempo reale. Per affrontare queste sfide, questa dissertazione estende il formalismo dei Wave Digital Filters (WDFs) in una struttura unificata e differenziabile per la modellazione e l'implementazione digitale di LEMs. Nel contesto dei WDFs, gli elementi circuitali e le reti di connessione topologiche sono descritti mediante blocchi ingresso-uscita caratterizzati da relazioni di scattering, in cui tensioni e correnti di porta sono espresse come combinazioni lineari di onde incidenti e riflesse grazie all'introduzione di una resistenza di riferimento per porta. La scelta di tale parametro libero per ciascuna porta offre importanti vantaggi numerici, che garantiscono stabilità, modularità ed efficienza computazionale. Sfruttando questa struttura modulare, la tesi introduce blocchi Wave Digital (WD) basati su reti neurali per rappresentare le equazioni di scattering nonlineari di elementi monoporta e multiporta. Sono considerate sia strutture WD ibride esplicite sia implicite: le prime consentono la simulazione in tempo reale di circuiti contenenti un singolo elemento WD neurale nonlineare, mentre le seconde permettono di trattare circuiti più complessi con molteplici blocchi WD nonlineari per mezzo di risolutori iterativi. Sia le strutture WD standard che quelle ibride possono essere simulate nel dominio tempo-discreto all'interno di strutture che implementino il paradigma dell'Automatic Differentiation (AD), consentendo il calcolo di derivate esatte della risposta simulata rispetto ai parametri del modello. Ciò rende possibile l’ottimizzazione basata sul gradiente sia dei parametri circuitali sia dei modelli neurali parametrici direttamente a partire da misure sperimentali. In questo scenario, i parametri dei LEM lineari e nonlineari possono essere affinati secondo un paradigma di apprendimento supervisionato, con blocchi WD neurali addestrati all'interno dello stesso ciclo di simulazione differenziabile. La struttura di modellazione proposta trova applicazione in due ambiti principali. Nel VA modeling, dove le strutture WDF ibride vengono utilizzate per emulare circuiti audio nonlineari con elevata accuratezza e capacità di esecuzione in tempo reale. Nella modellazione di altoparlanti MEMS piezoelettrici, dove LEM differenziabili che combinano implementazioni basate su WDF e blocchi di elaborazione basati su reti neurali vengono impiegati per descrivere il comportamento elettro-meccano-acustico nonlineare di tali trasduttori, supportati da procedure di identificazione dei parametri basate sui dati sperimentali acustici. Infine, estendendo l'approccio alla modellazione basato su LEMs, la tesi affronta l'inversione circuitale e la virtualizzazione di trasduttori introducendo un metodo sistematico per derivare circuiti inversi basato sui nullori. Questo permette la progettazione di algoritmi DSP in grado di compensare o modificare digitalmente il comportamento dei trasduttori audio tramite implementazioni digitali di sistemi diretti ed inversi opportunamente concatenati. L'approccio viene applicato ad un caso di studio sull'equalizzazione di un altoparlante MEMS piezoelettrico, ottenendo un appiattimento accurato della risposta in frequenza e preservando al contempo l'interpretabilità del modello fisico sottostante.
Differentiable circuital models of audio systems for digital signal processing
Massi, Oliviero
2025/2026
Abstract
Audio systems, ranging from analog effects and guitar amplifiers to modern MEMS loudspeakers, are inherently multiphysics and nonlinear devices. The nonlinear interactions among different physical domains, together with phenomena such as magnetic or piezoelectric hysteresis, are key to the distinctive behavior and sonic character of these systems. Capturing these effects in a lightweight digital representation is essential for both Virtual Analog (VA) modeling, which aims to faithfully emulate analog audio circuits in the digital domain, and for the design and optimization of audio transducers, where accurate physical models support the development of Digital Signal Processing (DSP) algorithms for performance enhancement. Lumped-Element Models (LEMs) provide a compact and physically interpretable framework for describing such systems by exploiting analogies between physical domains and representing them as equivalent electrical networks of interconnected circuit elements. These models are particularly well suited to audio applications, as they can be implemented efficiently in the discrete-time domain and interfaced with DSP structures. However, LEMs have limited expressive power when it comes to capturing strongly nonlinear or hysteretic effects, and their analytical treatment often leads to implicit equations that are difficult to solve in real-time. To address these challenges, this dissertation extends the Wave Digital Filter (WDF) formalism into a unified, differentiable modeling and implementation framework. In this approach, circuit elements and connection networks are described as input-output blocks characterized by scattering relations, where port voltages and currents are expressed as linear combinations of incident and reflected waves through the introduction of a reference port resistance. The choice of this free parameter per port provides important numerical advantages, ensuring stability, modularity, and computational efficiency. Building on this modular structure, in this thesis we introduce neural network-based Wave Digital (WD) blocks for representing the nonlinear scattering equations of one-port and multi-port elements. Both explicit and implicit hybrid WD structures are investigated: explicit realizations enable real-time simulation of circuits containing a single nonlinear neural element, while implicit realizations address more complex circuits with multiple nonlinear WD blocks using iterative solvers. Both standard and hybrid WD structures can be simulated in the discrete-time domain within an Automatic Differentiation (AD)-based framework, which enables the computation of exact derivatives of the simulated response with respect to the model parameters. This allows the gradient-based optimization of both circuit parameters and neural parametric models directly from experimental measurements. In this setting, the parameters of linear and nonlinear LEMs can be refined according to a supervised learning paradigm, with neural WD blocks trained within the same differentiable simulation loop. The proposed modeling framework is applied to two main areas. In VA modeling, hybrid WDF structures are used to emulate nonlinear audio circuits achieving high accuracy and real-time efficiency; in piezoelectric MEMS loudspeaker modeling, differentiable LEMs combining WDF implementations and neural network-based processing blocks are used to capture the nonlinear electro-mechano-acoustic behavior of these transducers, supported by data-driven parameter identification directly from acoustic measurements. Finally, building upon the established LEM modeling framework, we address circuit inversion and transducer virtualization, introducing a systematic method for deriving inverse circuit representations based on nullors. This enables the design of model-based DSP algorithms that digitally compensate or modify the behavior of audio transducers through cascaded direct and inverse systems. The approach is applied to a piezoelectric MEMS loudspeaker equalization case study, demonstrating accurate response flattening while preserving the interpretability of the underlying physical model.| File | Dimensione | Formato | |
|---|---|---|---|
|
Oliviero_Massi_PhD_Thesis.pdf
non accessibile
Descrizione: PhD Thesis in Information Technology
Dimensione
35.68 MB
Formato
Adobe PDF
|
35.68 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/256917