Speaker
Description
This paper studies learning-based sound field control for a prototype large-audience scenario with ten loudspeakers and multiple listeners distributed over a listening region. Two identical complex-valued multilayer perceptrons are compared. One is trained on a boundary normal-velocity objective and one is trained directly on binaural ear pressures. The velocity-based network is first validated against the analytical controller based on previous work by Shin et al and shown to closely reproduce its region-based behavior. The two models are then compared using common region-based and listener-based metrics, including pressure error and binaural cue errors based on Interaural Level Difference (ILD) and Interaural Phase Difference (IPD). For ten listeners, direct ear pressure training defines an overdetermined 20-ear/10-loudspeaker problem, whereas the velocity-based model provides a more robust trade-off between regional accuracy and perceptual performance. The results indicate that region-based velocity training is a promising approach for overdetermined multi-listener sound field reproduction.