An individualization approach for head-related transfer function in arbitrary directions based on deep learning

Abstract

This paper provides an individualization approach for head-related transfer function (HRTF) in arbitrary directions based on deep learning by utilizing dual-autoencoder architecture to establish the relationship between HRTF magnitude spectrum and arbitrarily given direction and anthropometric parameters. In this architecture, one variational autoencoder (VAE) is utilized to extract interpretable and exploitable features of full-space HRTF spectra, while another autoencoder (AE) is employed for feature embedding of corresponding directions and anthropometric parameters. A deep neural networks model is finally trained to establish the relationship between these representative features. Experimental results show that the proposed method outperforms state-of-the-art methods in terms of spectral distortion.