Deep Neural Network for Personalization of Parametric Head-Related Transfer Functions in a Median Plane

Head-related transfer functions (HRTFs) characterize how the human head and body modify the frequencies of sound waves as they travel toward the ear, thus aiding people in determining the direction and location of sound sources. HRTFs have different shapes that depend on the individual listener. Some studies have therefore used deep neural network (DNN) models to synthesize personalized HRTFs by measuring the sizes of the listener’s ears and head, but they have struggled to estimate large numbers of outputs (e.g., 200 samples in an impulse response or 512 samples in an HRTF). Therefore, in this work, we introduce parametric HRTF synthesis to reduce the number of outputs, and the DNN model synthesizes the HRTF in a median plane, in which the individual differences are likely to occur. The measured HRTF was approximated via series synthesis of six peaking digital filters, which were characterized in terms of center frequency, gain and bandwidth, and the output dimensions could then be reduced to 18 parameters. Use of this data compression process allowed the log spectral distance from the measured HRTF to be improved by 1 to $2 \%$, and psycho-acoustic experiments showed even difficulty to localize accurate in the $\mathbf{0}$ (front) and $\mathbf{1 8 0}$ (back) degree using the estimated parametric HRTF.