Audio spatialisation is a crucial component of immersive technologies, shaping the quality of experiences in VR, AR and gaming. Spatial audio rendering relies on Head-Related Transfer Functions (HRTFs), which encode both spatial localisation cues and listener-dependent spectral characteristics associated with perceptual attributes such as sound colouration. However, research has largely prioritised localisation accuracy, leaving the remaining subjectspecific spectral features underexplored. Meanwhile, 3D audio renderers often employ HRTFs with entangled spectral colouration profiles and localisation cues, leading to cases where localisation is accurate, but the resulting sound coloration is perceived as unnatural or undesirable. In this paper, we introduce a novel signal processing method, which we term the Delta Spatial Transfer Function (DSTF) which aims to separate low-order spectral characteristics from directional cues. We also present a conditional autoencoder model trained to reconstruct and synthesise new DSTF profiles. To ensure disentanglement between localisation and DSTF characteristics, we incorporate adversarial learning. The resulting model enables (1) synthesis of novel location-agnostic spectral characteristics, (2) neural compression of databases of such features, and (3) analysis and exploration of HRTF variation beyond spatial encoding.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=22993
(882KB)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
