ITD Modeling Based on Anthropometrics and KEMAR Coefficients Using Deep Neural Networks

Abstract

ITD and ILD, versus source arrival direction, serve as essential binaural cues for spatial hearing. Individualized ITD and ILD can be used to render better 3D audio than a non-individualized one. Due to the correlation between ITD and some anthropometric features, machine learning, such as principal component analysis (PCA) and deep neural networks (DNNs), have become important methods to deploy individualized ITDs. The available measured ITDs do not match the exact sound source directions. An ITD correction method will be presented to overcome the irregularities that occurr due to subject head movements during database creation measurements. KEMAR’s ITD coefficients are utilized to correct the misplacement of a subject’s ITD. DNNs are used to obtain a new subject’s ITD for 1250 different azimuth and elevation angles. Mean absolute error (MAE) is used to compare the proposed ITD model with the available analytical models.