A spherical basis function neural network for pole-zero modeling of head-related transfer functions

Abstract

This paper describes a neural network for approximating the parameters of a pole-zero model of the head-related transfer function (HRTF). The von Mises basis function (VMBF) is described whose response depends on spherical rather than Cartesian input coordinates. The VMBF neural network is ideally suited to the problem of learning a continuous mapping from spherical coordinates to acoustic parameters that specify sound source direction. A method for computing the common poles of a set of HRTFs is also discussed.