HRTF Individualization Based on Anthropometric Measurements Extracted from 3D Head Meshes

Abstract

In the field of 3D audio, the use of Head-Related Transfer Functions (HRTFs) compliant to the subject anatomical traits is crucial to guarantee a proper individual experience. This work proposes an HRTF individualization method based on anthropometric features automatically extracted from 3D head meshes. The method aims at a fully automated process able to estimate individual median plane HRTF starting from a 3D mesh of the subject’s pinna. The method relies on the HUTUBS dataset including 3D meshes, anthropometry and HRTFs. In the first phase, a set of pinna anthropometric parameters is extracted from the 3D meshes converted to range images. A set of landmarks is fitted on the pinna through the Active Shape Model algorithm to outline its shape. Then, the set of pinna anthropometric parameters defined in HUTUBS is automatically extracted exploiting the landmarks. In the second phase, the relationship between pinna anthropometry and HRTFs is modelled. For each elevation angle considered in HUTUBS, a Generalized Regression Neural Network is trained to predict the corresponding HRTF, given the anthropometry. The method is evaluated in both objective and perceptual metrics showing performances comparable to the state of the art.