Pressure-matching (PM) for personal sound zone (PSZ) can achieve high contrast at nominal control points, but the performance may degrade when transfer functions are mismatched. We introduce a neural method that maps transfer functions to loudspeaker weights using a single-frequency input network with parameters shared across frequencies. We evaluate the robustness under position shifts, additive transfer-function noise, and added reflections, and compare against PM with Tikhonov regularization. Results show improved robustness to structured perturbations such as listener displacement, whereas regularized PM remains more resilient to unstructured random transfer-function noise and reverberation. We further explain these results using a singular value decomposition based perturbation projection. Finally, we analyze different regularization mechanisms induced by the network and derive practical guidelines for neural PSZ filter optimization.
