Speaker
Description
The standard image source method approximates surface reflections using plane wave reflection coefficients, which can introduce errors at low frequencies. The complex image source method addresses this issue by using reflection coefficients derived from the Sommerfeld integral solution for a point source above a sound-absorbing plane of infinite extent. However, this approach is more computationally expensive than the standard image source method because the spherical wave reflection coefficient must be computed for each image source, which limits its practical applications. In this study, we propose a multilayer perceptron that models the reflection of spherical waves at the surfaces of a reverberant room. This surrogate model is trained on solutions to the Sommerfeld integral. In an image source implementation, for each image, the surrogate model receives the distance between the image and the receiver, and the angle of reflection. The model outputs a corresponding spherical wave reflection coefficient. The resulting modified image source method is more accurate than the standard and complex image source methods when compared to finite element solutions of the room transfer functions.