Hear-to-Hear: Uncertainty-Aware, Head-Tracked Interaction for Personalized Spatial Audio

Abstract

Selecting a personalized head-related transfer function (HRTF) is essential for accurate spatial audio, but remains challenging due to substantial user variation and noisy subjective feedback. We present a real-time, uncertainty-aware perceptual interaction loop for HRTF personalization: after each paired comparison, the system updates a posterior belief over candidate HRTFs and immediately adapts the next query to maximize informativeness. Our system couples Bayesian preference learning with real-time head-tracked binaural rendering, enabling head motion to actively inform comparisons. In a controlled within-subject psychophysical study, uncertainty-aware interaction (with optional head tracking) provides directional posterior evidence for localization improvements－primarily suggestive reductions in azimuth error and front-back confusions concentrated in the front hemisphere－alongside participant-level heterogeneity and a partially mitigated elevation trade-off, rather than uniformly decisive effects across all metrics.