Perception-informed alignment of learning personalized head-related transfer function latent representations

Computational auditory perception has been developed and recently used to validate machine learning systems prior to subjective listening tests. An emerging challenge is integrating perceptual science into machine learning model training to improve perceptual modeling. In this work, we focus on personalized head-related transfer function (HRTF) representation learning, incorporating perceptual constraints into deep learning. We first study whether latent representations learned from existing deep learning methods preserve perceptual relations using Pearson correlation between latent space distances and objective perceptual metrics. We then introduce an implicit neural representation (INR)-based approach that embeds HRTFs into a perception-informed latent space, ensuring the learned representations are not only spectrally accurate but also perceptually relevant. This method combines a metric-based loss function and metric multidimensional scaling (MMDS) supervision, even when objective perceptual metrics are non-differentiable. Our approach significantly enhances correlations between learned representations and perceptual dimensions, including localization, coloration, and externalization. We demonstrate that these representations can enhance personalized HRTF selection, suggesting the potential to improve listening experiences. While subjective listening remains the gold standard, our method provides a stronger integration of perceptual science into model development, facilitating perceptually aligned spatial audio rendering.