Median-Plane Head-Related Transfer Function Personalization Using Two-Dimensional Independent Component Analysis

A new median-plane head-related transfer function (HRTF) personalization method using independent component analysis and ensemble learning is proposed. First, an HRTF personalization model based on two-dimensional independent component analysis is constructed in which HRTFs are decomposed as a combination of independent components from both the frequency-direction and subject domains. Then, for a new subject, subject-related independent components are predicted by substituting the subject's anthropometric parameters into an ensemble learning model consisting of three generalized regression neural networks (GRNNs), namely, a head-GRNN, torso-GRNN, and pinna-GRNN. Finally, the subject's individual HRTFs are predicted by substituting the predicted subject- related independent components into the HRTF personalization model. Results suggest that the proposed method can be used to personalize HRTFs effectively among direction, frequency, and subject, yielding lower spectral distortion values for both ears when compared to other similar methods.