Speaker
Description
Accurately evaluating sound localisation performance is essential for developing and validating methods to personalise head-related transfer functions (HRTFs). Localisation in the median plane is typically summarised with complementary metrics: polar error for local precision, polar gain for spectral decoding ability, and quadrant error rate for front-back confusions. These metrics rely on hard thresholds applied to subsets of trials, which limits statistical robustness and compresses individual differences (for instance, polar error excludes responses beyond 90° of the target, capping polar error at 52° for listeners responding at chance). Here, we propose a probabilistic model of the complete polar response distribution, based on a four-component von Mises mixture, that requires no data partitioning and negligible computation. The model free parameters are jointly identifiable, as confirmed by a synthetic recovery study, and can be estimated from as few as 45 trials per listener. Fitted to 33 listeners, each parameter correlates with a classical metric: confusion rate onto quadrant error rate, concentration onto polar error, and a spatial prior parameter onto polar gain. Model comparison further supports including the spatial prior for the large majority of listeners. The framework lends itself naturally to Bayesian extensions, offering a principled basis for more data-efficient individualised HRTF evaluation.