Perceptual Impact on Localization Quality Evaluations of Common Pre-Processing for Non-Individual Head-Related Transfer Functions

Abstract

This article investigates the impact of two commonly used Head-Related Transfer Function (HRTF) processing/modeling methods on the perceived spatial accuracy of binaural data by monitoring changes in user ratings of non-individualized HRTFs. The evaluated techniques are minimum-phase approximation and Infinite-Impulse Response (IIR) modeling. The study is based on the hypothesis that user-assessments should remain roughly unchanged, as long as the range of signal variations between processed and unprocessed (reference) HRTFs lies within ranges previously reported as perceptually insignificant. Objective assessments of the degree of spectral variations between reference and processed data, computed using the Spectral Distortion metric, showed no evident perceptually relevant variations in the minimum-phase data and spectral differences marginally exceeding the established thresholds for the IIR data, implying perceptual equivalence of spatial impression in the tested corpus. Nevertheless analysis of user responses in the perceptual study strongly indicated that variations introduced in the data by the tested methods of HRTF processing can lead to inversions in quality assessment, resulting in the perceptual rejection of HRTFs that were previously characterized in the ratings as the "most appropriate" or alternatively in the preference of datasets that were previously dismissed as "unfit." The effect appears more apparent for IIR processing and is equally evident across the evaluated horizontal and median planes.

Spatial up-sampling of HRTF sets using generative adversarial networks: A pilot study

“…The HRTF phase was disregarded, assuming that the up-sampled HRTFs would be reconstructed using a minimum-phase approximation and a simple ITD model. It is known that such simplifications could have an impact on certain perceptual features of the HRTFs Andreopoulou and Katz, (2022), therefore further research beyond this pilot will probably need to also consider phase information.…”

Initial Evaluation of an Auditory-Model-Aided Selection Procedure for Non- Individual HRTFs

“…Worst 0 training on the VR pointing procedure (similar to [17]) and investigate HRTF processing algorithms [26]. Finally, a high level of front-back reversal might have been caused by the visual environment dominating the perception when the non-individual auditory cues were vague: lack of visual source in the front of the listener led participants to believe that the sound must have come from their back (even if the subjects were informed that the sounds might come from any direction, both front and back).…”

Optimization of piano tuning by means of spectral entropy minimization

“…Even though time is more intuitive than frequency to be modeled as the fourth dimension, time-based approach to HRTF modeling has been proven to be less accurate than the frequency-based one [20,35]. Furthermore, HRIRs include information on phase, which is widely acknowledged to be irrelevant to localization as long as ITD is preserved [8,36,37], although conflicting results have been presented regarding whether or not the phase linearization is detectable, particularly at low frequencies [16,38,39]. This work focuses only on the magnitude part of HRIR spectra, assuming that the phase can be either linearized basing on ITD or modeled independently.…”

“…Breebaart et al investigated parametrizing HRTFs on ERB scale, where for most positions around 20 parameters were needed to achieve perceptual irrelevance of the approximation [47]. On the other hand, Andreopoulou and Katz reported results, in which minimum-phase HRTFs for η = 32 were clearly discernible from full-phase 256-sample data [39]. This discrepancy might be caused by using prescreening, which ensured that the subjects were very competent for the task, with above-average hearing abilities.…”