A Bayesian framework for the estimation of head-related transfer functions

Abstract

While high-fidelity spatial auditory displays require individualized head-related transfer functions (HRTFs), much of the physical structure contained within an HRTF is similar across most individuals. This suggests that a Bayesian estimation technique, which uses sample observations to bias an a priori model towards an individual, may provide benefits in terms of efficiency. Therefore, the current work proposes a Bayesian HRTF framework that utilizes HRTFs in the form of their real spherical harmonic representation. When combined with assumptions of normality, the resulting technique is shown to enable the accurate estimation of an individualized HRTF from a small set of spatially distributed measurements. Moreover, the model provides a convenient way to analyze which components of an HRTF vary most across individuals, and can therefore be used to create very efficient HRTF measurement strategies. A perceptual localization test confirmed that similar localization performance could be attained with an HRTF estimated from as few as 12 spatial measurements, even when confined to the median-sagittal plane.