HRTF Recommendation Based on the Predicted Binaural Colouration Model

Over the past few decades, several databases of head-related transfer functions (HRTFs) have been recorded using both mannequin heads and individual subjects. Some of these databases also include the measurements of the subjects’ anthropometric features (e.g. pinna height, head height, neck width etc). This paper is concerned with the task of recommending one HRTF set out of an entire database based on the listener’s own anthropometric features alone. Furthermore, it focuses specifically on minimising the colouration that the listener would perceive when using a given HRTF set compared to their own individual one. In contrast to prior work, the proposed methods leverage an objective measure, called Predicted Binaural Colouration (PBC), which has been previously shown to correlate strongly with perceived binaural colouration. This measure facilitates the generation of large amounts of data, which in turn enables training larger models than previously possible. Both a random forest regressor and a neural network are trained, the latter accounting for 74.2% of the colouration variance between HRTFs from morphological differences. This highlights the potential utility of the presented approach in recommendation systems. The study further identifies the intertragal incisure width asymmetry as the pinnae-related feature most correlated with PBC.