Virtual autoencoder based recommendation system for individualizing head-related transfer functions

We propose a virtual autoencoder based recommendation system for learning a user´s Head-related Transfer Functions (HRTFs) without subjecting a listener to impulse response or anthropometric measurements. Autoencoder neural-networks generalize principal component analysis (PCA) and learn non-linear feature spaces that supports both out-of-sample embedding and reconstruction; this may be applied to developing a more expressive low-dimensional HRTF representation. One application is to individualize HRTFs by tuning along the autoencoder feature spaces. We demonstrate this new approach by developing a virtual (black-box) user that can localize sound from query HRTFs reconstructed from those spaces. Standard optimization methods tune the autoencoder features based on the virtual user´s feedback. Experiments with CIPIC HRTFs show that the virtual user can localize along out-of-sample directions and that optimization in the autoencoder feature space improves upon initial non-individualized HRTFs. Other applications of the representation are also discussed.