On the 1st of July, Judy NAJNUDEL in the S3AM team, has defended her thesis called:
Power-Balanced Modeling of Nonlinear Electronic Components and Circuits for Audio Effects
Her thesis is directed by Thomas HELIE, CNRS, head of the S3AM team, STMS (Ircam, Sorbonne Université, CNRS, ministère de la Culture) and co-directed by David ROZE (CNRS-STMS) and Rémy MÜLLER (UVI).
It was at IRCAM, Stravinsky room, but you can follow it through the IRCAM's YouTube channel: https://youtu.be/v3tq6A--WuE
Jury:
Bernhard Maschke - Rapporteur - Professeur, Université Lyon 1
Udo Zölzer - Rapporteur - Professeur, Université Helmut Schmidt de Hambourg
Benoît Fabre - Examinateur - Professeur, Sorbonne Université
Antoine Falaize - Examinateur - Ingénieur de recherche, Université de La Rochelle
Stefania Serafin - Examinatrice - Professeure, Université de Aalborg
Thomas Hélie - Directeur de thèse - DR CNRS, STMS
David Roze - Co-encadrant - CR CNRS, STMS
with the guests: Rémy Müller (UVI) et Manuel Schaller (TU Ilmenau)
ABSTRACT:
This thesis is concerned with the modeling of nonlinear components and circuits for simulations in audio applications. Our goal is to propose models that are sufficiently sophisticated for simulations to sound realistic, but that remain simple enough for real time to be attainable.
To this end, we explore two approaches, both based on a port-Hamiltonian systems formulation. Indeed, this formulation structurally guarantees power balance and passivity. Combined with ad hoc numerical methods, this ensures the numerical stability of simulations.
The first approach is comparable to "white box" modeling. It assumes that the circuit topology is known, and focuses on the modeling of specific components found in vintage audio circuits, namely ferromagnetic coils (found in wah-wah pedals and guitar amplifiers) and opto-isolators (found in tremolos and optical compressors). The proposed models are physically-based, passive, modular, and usable in real time.
The second approach is comparable to "grey box" modeling. It aims to retrieve the topology and constitutive laws of a circuit from measurements. The learning of the circuit topology is informed by an underlying port-Hamiltonian formulation, and nonlinearities are concomitantly addressed through kernel-based methods. Thus, necessary physical properties are enforced, while the use of reproducing kernels allows for a variety of
nonlinear behaviors to be described with a smaller number of parameters and a higher interpretability compared to neural network methods. Finally, a possible generalization of this approach for a larger class of circuits is outlined through the introduction of the Koopman operator.