Virtual, augmented, and mixed reality experiences are becoming more commonplace as consumer-grade devices proliferate. Head-Related Transfer Functions (HRTFs) are used to create realistic spatial audio in virtual and augmented environments. Mathematically, HRTFs represent solutions to acoustic boundary-value scattering problems governed by the Helmholtz equation. Neural operators are neural networks designed to learn the solutions of partial differential equations (PDEs). The present work proposes an operator-learning framework based on the Deep Operator Network (DeepONet) for individualized HRTF prediction. By implementing a non-uniform sampling strategy for 3-D head meshes and data compression along the frequency axis, the framework achieves high-fidelity predictions while reducing data dimensionality. Our method shows low log-spectral distortion, generalizes to unseen spatial grids, and infers an entire head’s HRTF field in ~0.3 seconds. Objective evaluations demonstrate the framework's effectiveness in personalization and spatial interpolation. Furthermore, robust performance on unseen subjects and coordinates highlights the model's generalization capability, offering a computationally efficient alternative for HRTFs personalization.
