Tuning of head-related transfer functions using principal component analysis.

Abstract

Modeling head related impulse responses (HRIRs) to determine head-related transfer functions (HRTFs) can be used to create a virtual auditory display through head phones. However, non-individualized HRTFs can create errors in sound localization and perception of the source as coming from inside vs outside of the head. We present a method, based on model-order reduction and principal component analysis, to “tune” generalized HRTFs derived from the CIPIC database of HRIRs in order to individualize an HRTF. The impulse response for each individual in the CIPIC database is modeled at given azimuth locations on the horizontal plane. Model order reduction is used to eliminate poles and zeros of the identified HRTF at high frequencies, while preserving the frequency response below ∼8000 Hz. Principal component analysis is then performed on the reduced-order models in order to reduce the dimensionality of the dataset and identify a subset of weighting variables that are adjusted during HRTF customization. The sensitivity of weights on the frequency response is also studied in order to correlate weights with HRTF features. A prototype auditory display has been developed that allows a user to tune HRTFs in real-time, while listening to a 44.1-kHz source presented from a chosen azimuth.