HRTF selection for binaural synthesis from a database using morphological parameters

Auditory virtual environments are becoming increasingly relevant for applications such as teleconferencing, hearing aids, video games, and general immersive listening. To enable high fidelity renderings of the sound scene in such environments, the audio content must be treated with the actual listener's acoustical filters, the so-called head-related transfer functions (HRTFs). The current challenge for general public applications, given the difficulty of measuring HRTFs for a given listener, is to be able to individually generate HRTFs or perform a selection from a database of pre-existing HRTFs, so as to provide the listener an HRTF that enables a listening experience that is as realistic as possible, using for example only data taken from a photo of the listener's ear. A process is described in which a database of 46 measured HRTFs was analysed using various data reduction techniques such as principal component analysis and frequency scaling. A selection of the subjects' most significant morphological parameters was performed using data mining techniques such as support vector machines. This subset of morphological parameters for subjects associated to the HRTF database were then used to perform multiple linear regressions against the reduced dataset of HRTFs in order to predict what might be the listener's preferred HRTFs. The prediction performance was then compared to the results of a perceptual evaluation of the HRTFs from the database using a listening test. The results show that the proposed process was able to predict preferred HRTFs for a listener significantly better than if the HRTFs were chosen at random. The results from the listening test were also used to explore a perceptually relevant frequency range of the HRTF.