Head-Related Transfer Function Selection Using Neural Networks

Abstract

In binaural audio systems, for an optimal virtual acoustic space a set of head-related transfer functions (HRTFs) should be used that closely matches the listener's ones. This study aims to select the most appropriate HRTF dataset from a large database for users without the need for extensive listening tests. Currently, there is no way to reliably reduce the number of datasets to a smaller, more manageable number without risking discarding potentially good matches. A neural network that estimates the appropriateness of HRTF datasets based on input vectors of anthropometric measurements is proposed. The shapes and sizes of listeners' heads and pinnas were measured using digital photography; the measured anthropometric parameters form the feature vectors used by the neural network. A graphical user interface (GUI) was developed for participants to listen to music transformed using different HRTFs and to evaluate the fitness of each HRTF dataset. The listening scores recorded were the target outputs used to train the neural networks. The aim was to learn a mapping between anthropometric parameters and listener's perception scores. Experimental validations were performed on 30 subjects. It is demonstrated that the proposed system produces a much more reliable HRTF selection than previously used methods.

HRTF Individualization: A Survey

“…However, as reported in Table I, the perceptual results are mixed. On one side, the simple methods, namely selection and adaptation by frequency scaling and/or set rotation, have demonstrated some perceptual improvement compared to no individualization, thanks to studies that featured 6 to 11 participants [29,34]. On the other side, we cannot conclude on the quality of the HRTFs produced by more complex methods, such as linear and nonlinear regression between anthropometric measurements and HRTF sets.…”

“…For instance, using the CIPIC database [32], Zotkin [33] implemented in 2002 a coarse nearest neighbors approach that used only 7 morphological parameters measured on a picture of the pinna, and showed some improvement in terms of localization performance compared to no individualization (average gain of 15% in elevation score). More recently, in 2017, Yao [34] proposed a more exotic method to select a HRTF set among a database, using a neural network trained to predict a perceptual score (from 1 to 5) from anthropometric measurements. However, it is difficult to conclude on the results of their perceptual study in comparison with others, as it only used their own perceptual score as indicator.…”

Individualizing Head-Related Transfer Functions for Binaural Acoustic Applications

“…Nearest neighbor selection: In these approaches, the nearest HRTF set in a dataset is first selected based on the anthropometric measurements. The distances between two subjects can be computed either directly from morphological parameters [54], or features output from a neural network [53]. Adaption can be further applied using methods in the previous category.…”

HRTF Individualization using Deep Learning

“…In more recent times, there has been an interest in solving these tasks using deep learning techniques [7]. In 2017, Yao et al [29] used anthropometric measurements to select the most suitable HRTF sets from a larger database. In their work, a dataset of user anthropometry and fitness scores for each available HRTF is compiled -by means of conducting perceptual tests with users -and neural networks for each HRTF are trained to predict their suitability.…”