Head-related transfer functions upsampling with physics-informed spherical convolutional neural network

Abstract

Head-related transfer functions (HRTFs) play a crucial role in virtual acoustics and spatial audio applications. However, obtaining personalized, high-resolution HRTFs remains challenging due to the time-consuming and expensive measurements. Recently, deep learning-based methods have shown promise in predicting high-resolution HRTFs from sparse measurements. Nevertheless, most of these methods often treat HRTF upsampling as an image super-resolution task, which overlooks critical spatial information and acoustic principles, leading to overfitting on datasets. This paper proposes a physics-informed spherical convolutional neural network for HRTF upsampling. First, spherical convolutional layers are used to capture spatial features on the sphere, allowing efficient handling of spherical sampled HRTF data. Second, in the upsampling process, the proposed method incorporates the Helmholtz equation as a constraint, adhering tothe physics of the acoustic system. This method ensures the generation of physically feasible HRTF interpolations, thus promoting better generalization.