Communication Dans Un Congrès Année : 2025

Data-driven Modeling of Musical Instruments: First Attempts

Résumé

The traditional approach to modeling physical phenomena relies on applying first principles to derive a set of equations. The equations are then employed to generate numerical data. Finally, this data is compared with experimental results to validate the model. An alternative approach has emerged recently: constructing a model directly based on experimental data generated by the phenomenon under study. According to this data-driven approach, the model consists of the set of equations that best explains the data. In the literature, various methods have been proposed to identify these optimal equations, ranging from symbolic computation to the use of neural networks. Our focus is on another approach, where equations are built by combining basis functions available in a user-defined library. A sparsity constraint is added to the process to derive equations that contain the minimal number of elements from the function library. This approach is formulated as a regression problem (minimizing the L2 norm of the model's deviation from experimental data) with regularization to promote sparsity (minimizing the L1 norm of the activation matrix of the library's functions). The method has been tested on data initially generated by models of self-oscillating systems and subsequently on experimental results. We will demonstrate how this method can also propose models that account for the continuous evolution of a parameter during the experiment.

Fichier principal
Vignette du fichier
Présentation1.pdf (8.01 Mo) Télécharger le fichier

Dates et versions

hal-05365562 , version 1 (17-04-2026)

Licence

Identifiants

  • HAL Id : hal-05365562 , version 1

Citer

Christophe Vergez, Vincent Freour, Aziz Chedhli, Bruno Cochelin, Keita Arimoto. Data-driven Modeling of Musical Instruments: First Attempts. CFA 2025 - 17e Congrès Français d'Acoustique, Société Française d'Acoustique (SFA), Apr 2025, Paris, France. ⟨hal-05365562⟩
31 Consultations
3 Téléchargements

Partager

  • More