IT'S A DIVE

The article “HRTF selection by anthropometric regression for improving horizontal localization accuracy” (S. Spagnol) was recently accepted for publication in the IEEE Signal Processing Letters. An early access version of the article is currently available at the following URL:
https://ieeexplore.ieee.org/document/9050904. It will be made open access in its final version.

The article focuses on objective HRTF selection from anthropometric measurements for minimizing localization error in the frontal half of the horizontal plane. Localization predictions for every pair of 90 subjects in the HUTUBS database are first computed through an interaural time difference-based auditory model, and an error metric based on the predicted lateral error is derived. A multiple stepwise linear regression model for predicting error from inter-subject anthropometric differences is then built on a subset of subjects and evaluated on a complementary test set. Results show that by using just three anthropometric parameters of the head and torso (head width, head depth, and shoulder circumference) the model is able to identify non-individual HRTFs whose predicted horizontal localization error generally lies below the localization blur. When using a lower number of anthropometric parameters, this result is not guaranteed.

The new year kicks off with two paper acceptances. “Auditory model based subsetting of head-related transfer function datasets” (S. Spagnol) will be presented at the 45th IEEE International Conference on Acoustics, Speech, and Signal Processing (IEEE ICASSP 2020) taking place in Barcelona, Spain in May. We will also be in Atlanta, US next month at the IEEE 5th VR Workshop on Sonic Interactions in Virtual Environments with the paper “HRTF individualization using deep learning” (R. Miccini, S. Spagnol), as part of the 27th IEEE Conference on Virtual Reality and 3D User Interfaces (IEEE VR 2020).

The former paper outlines a novel HRTF subset selection algorithm based on auditory-model vertical localization predictions and a greedy heuristic to identify an optimal subset of representative HRTFs from a catalogue including the three biggest open HRTF datasets currently available online (see figure above). An objective validation of the optimal subset on a fourth independent dataset, based again on auditory model predictions, is provided. The results show an overwhelming agreement on the choice of this optimal subset, supporting the idea that a large HRTF catalogue can be efficiently reduced by two orders of magnitude while preserving at least one HRTF fitting the very large majority of a pool of listeners in terms of localization error.

The research presented in the latter paper focuses on HRTF individualization using deep learning techniques. The rising availability of public HRTF data currently allows experimentation with different input data formats and various computational models. Accordingly, three research directions are investigated: (1) extraction of predictors from user data (see figure above) ; (2) unsupervised learning of HRTFs based on autoencoder networks; and (3) synthesis of HRTFs from anthropometric data using deep multilayer perceptrons. While none of the aforementioned investigations has shown outstanding results to date, the knowledge acquired throughout the development and troubleshooting phases highlights areas of improvement which are expected to pave the way to more accurate models for HRTF individualization.

Our paper “Estimation of pinna notch frequency from anthropometry: an improved linear model based on Principal Component Analysis and feature selection” (R. Miccini, S. Spagnol) has been accepted for presentation at the 1st Nordic Sound and Music Computing Conference that will take place in Stockholm, Sweden next month.

In the paper, anthropometric data from a database of HRTFs is used to estimate the frequency of the first pinna notch in the frontal part of the median plane. Given the presence of high correlations between some of the anthropometric features, as well as repeated values for the same subject observations, we propose the introduction of Principal Component Analysis (PCA) to project the features onto a space where they are more separated. We then construct a regression model employing forward step-wise feature selection to choose the principal components most capable of predicting notch frequencies. Our results show that by using a linear regression model with as few as three principal components (M₁ in the above plot), we can predict notch frequencies with a cross-validation mean absolute error of just about 600 Hz.