HRTF Individualization using Compact PRTF for Accurate Spectral Cues

Spatial audio rendering relies on accurate localization perception, which necessitates individualized head-related transfer functions (HRTFs). Previous deep neural network (DNN) approaches for predicting HRTF magnitude spectra from pinna images primarily used HRTF log-magnitude as the training output. However, because HRTFs encompass the acoustic properties of the head and torso, reconstructing the spectral cues essential for elevation localization remains a significant challenge. To address this issue, we introduce PRTFNet, an end-to-end convolutional neural network (CNN) designed to reconstruct individual spectral cues in HRTFs while minimizing the influence of head and torso reflections. PRTFNet utilizes the compact pinna-related transfer function (PRTF) as its output, which isolates the pinna’s contribution by excluding head and torso reflections from the head-related impulse response (HRIR). Evaluation using the HUTUBS dataset demonstrates that PRTFNet effectively reconstructs critical spectral cues, such as the first and second spectral notches. It surpasses previous deep learning models in performance, achieving lower log spectral distortion (LSD) and effective LSD (LSD E ), thereby enhancing the precision of spectral cue reconstruction for spatial audio rendering.