A Process for Calibrating HRTFs Based on Differentiable Implicit Representations and Domain Adversarial Learning

Individualizing Head Related Transfer Functions (HRTFs) can enhance users experience in settings like virtual reality, gaming and spatial music. Normally, individualization approaches are based on the anthropometric measurements of people, either by using a nearest-neighbor approach on a dataset or by regressing those measurements to individualized HRTFs. Those approaches, however, are either too time intensive and/or produce HRTFs sets “as is”, not allowing the user to fine-tune the resulting HRTFs. We propose in this paper a process in which a short listening test (of a couple of minutes) could be performed so that a start HRTFs set is calibrated in order to improve the accuracy of the perceived sound localization. This is done by optimizing parameters of an implicit HRTF representation with a differentiable loss. Additionally, we show that we can improve learned HRTFs representation by leveraging a domain adversarial loss during training. We present results in simulated participants, showing that the method works in theory and provides equivalent results as an optimal nearest-neighboring approach on a large (1000+ HRTFs) dataset.