Article ID: e26.97
In metaverse and digital twin applications, accurate reproduction of sound source radiation characteristics is essential to deliver a realistic and immersive acoustic experience. While accurate reconstruction generally necessitates dense spherical microphone arrays, these setups often prove unfeasible because of their disruption to the sound field and the complexity of their deployment. To address this issue, spatial signal reconstruction techniques are required, which can recover the radiation characteristics from sparsely sampled measurements. However, conventional methods based on local interpolation or global basis expansion have limitations in reconstructing sharp spatial variations and extrapolation regions under sparse and irregular sampling. In this paper, we propose and compare two Graph Signal Processing (GSP)-based methods for frequency-domain estimation of source radiation characteristics: a geometry-based graph method (GSP-Geo) and a spherical-harmonic-based graph method (GSP-SH). For the loudspeaker and reference-direction configuration examined in this study, the two GSP-based methods exhibited complementary behavior in the extrapolation region: GSP-Geo yielded lower errors at low and middle frequencies, whereas GSP-SH yielded lower errors at high frequencies.