End-to-End Learning for Physics-Based Acoustic Modeling

In past years, physics-based acoustic modeling developed theoretically to the point of yielding accurate understanding and description of a large number of acoustic phenomena, such as those involved in sound generation. Numerical algorithms have been proposed that are able to simulate these phenomena in real time with an acceptable computational cost, indeed reaching the market with commercial products. Sound synthesis based on physical models could benefit greatly from automated methods that require less specific know-how and save the sound-designer valuable time. This paper introduces a novel approach to parameter estimation in physics-based sound synthesis that is general and obtains good results based on an end-to-end computational intelligence paradigm. The approach is presented in a formal way and application to a practical use case is reported. Methodological issues, such as dataset generation, are investigated.

I. Introduction

Research in physics-based acoustic modeling obtained remarkable results in understanding and developing models for complex acoustic phenomena such as spatial sound diffusion, plate and membrane vibration, or string excitation, just to name a few. When the physical models are transformed in the discrete-time domain, numerical algorithms can be designed to simulate an acoustic phenomenon to the extent of obtaining a perceptually accurate rendition of a real-world sound. To obtain such a goal, a part of the effort must be devoted to the estimation of the model parameters. The more the model departs from its physical behavior, the more it gets harder to estimate the parameters, because they cannot be measured or inferred from physical properties. This is often the case with those techniques where some simplifying hypotheses are assumed to reduce the computational cost or the model complexity, such as Digital Waveguide modeling (DWG) [1]. Similarly, the more the model gets accurate, e.g. introducing secondary physical effects, the more the parameters increase.