M2 Internship – Machine learning and sound simulations by physical models to help the design of musical instruments

Sound simulations by physical models are interesting to transcribe the physics underlying the functioning of a musical instrument. These simulations make it possible to listen to a virtual instrument with a mode of operation representative of the interaction musician/instrument (driving a sound by the causes that create it). But the computation times of these simulations is heavy and only few designs are possible to represent the physical behavior of a musical instrument. The idea, implemented in 2 previous projects (2023 and 2024), is to train a machine learning model (ML) on a dataset of simulated designs. Random forest and elastic net regression were implemented with success in the two previous projects. The results show that the fitting of the ML models is satisfying and that accurate predictions of the behavior of instruments can be made on the whole design space. The application carried out concerns the brass instruments (trumpet, trombone), for which the input of the ML model is the geometry of the resonator and the outputs are descriptors characterizing the sound and playability (intonation and pressure threshold).
The main objective of this project is to improve the previous process in order to support the design of brass instruments. In particular, the goals will be:
– to analyze the ML models in order to understand the relationships between the input variables of the models and the sound descriptor. In this context, sensitivity analysis can be performed, to understand the effects of input variables on the sound descriptors. Different maps must be elaborated (using data analysis methods e.g. PCA) to represent these relationships and help the instrument design,
– to consider new descriptors of the behavior of instruments, related to the timbre of the instrument (e.g. spectral centroid), or the attack time, not considered in the previous internships, and to fit ML models on the simulated data,
– to confirm that the predictions, made on virtual prototypes, are valid when real prototypes are considered. For this, prototypes of a part of a resonator (typically the leadpipe of a trumpet) could be manufactured using rapid prototyping techniques, and tested with musicians, in collaboration with Yamaha corporation, who will provide different case studies and a realistic industrial context,
– to improve the sound simulation techniques, by considering an improved model of the excitator (the lips of the virtual musician), and by the integration of non-linear propagations in the resonator.

More information on the link below.