Cross-Dataset Head-Related Transfer Function Harmonization Based on Perceptually Relevant Loss Function

Head-Related Transfer Functions (HRTFs) play a vital role in binaural spatial audio rendering. With the release of numerous HRTF datasets in recent years, abundant data has become available to support HRTF-related research based on deep learning. However, measurement discrepancies across different datasets introduce significant variations in the data and directly merging these datasets may lead to systematic biases. The recent Listener Acoustic Personalization Challenge 2024 (European Signal Processing Conference) dealt with this issue, with the task of harmonizing different datasets to achieve lower classification accuracy while meeting thresholds over various localization metrics. To mitigate cross-dataset differences, this paper proposes a neural network-based HRTF harmonization approach aimed at eliminating dataset-specific properties embedded in the original measurements. The proposed method utilizes a perceptually relevant loss function, which jointly constrains multiple objectives, including interaural level differences, auditory-filter excitation patterns, and classification accuracy. Experimental results based on eight datasets demonstrate that the proposed approach can effectively minimize distributional disparities between datasets while mostly preserving localization performance. The classification accuracy for harmonized HRTFs between different datasets is reduced to as low as 31%, indicating a significant reduction in cross-dataset discrepancies. The proposed method ranked first in this challenge, which validates its effectiveness.