Spatial up-sampling of HRTF sets using generative adversarial networks: A pilot study

Headphones-based spatial audio simulations rely on Head-related Transfer Functions (HRTFs) in order to reconstruct the sound field at entrance of listener’s ears. A HRTF is strongly dependent specific anatomical structures, and it has been shown that virtual sounds recreated with someone else’s result worse localisation accuracy, as well altering other subjective measures such externalisation realism. Acoustic measurements filtering effects generated by ears, head torso proven be one most reliable ways obtain a personalised HRTF. However this requires dedicated expensive setup, time-intensive. In simplify measurement thereby improving scalability process, we are exploring strategies reduce number acoustic without degrading resolution Traditionally, up-sampling sets achieved through barycentric interpolation or employing spherical harmonics framework. However, methods often perform poorly when provided data spatially very sparse. This work investigates use generative adversarial networks (GANs) tackle problem, offering an initial insight about suitability technique. Numerical evaluations based spectral magnitude error perceptual model outputs presented single dimensions, therefore considering sources positioned only three main planes: Horizontal, median, frontal. Results suggest traditional better than proposed GAN-based distance between smaller 90°, but for sparsest conditions (i.e., every 120°–180°), approach outperforms others.