Magnitude Modelling of Individualized HRTFs Using DNN Based Spherical Harmonic Analysis

Data-driven individualized HRTF models, such as DNN-based models, strongly depend on the data while most existing HRTF databases have a limited number of measurements on listening subjects. In addition, different HRTF databases have different measurement positions and conditions, making the fusion of these databases challenging. This paper proposes a method for magnitude modelling of an individualized HRTF using DNN-based spherical harmonic analysis of the HRTF data. The HRTF log-magnitude spectra are decomposed using spherical harmonics, based on which multiple HRTF databases can be combined to effectively increase the amount of data for the follow-up DNN modelling. Then, multiple sub-networks are trained to map between the anthropometric parameters and decomposed spherical harmonic coefficients, based on which the individualized HRTF magnitude spectra are generated. Through experimental validation and comparison with the existing data-driven approaches, we show that the proposed method has more accurate modelling performance and lower modelling complexity. In addition, with predicted spherical harmonic coefficients, an individualized HRTF of arbitrary direction can be generated.