Perceptually enhanced spectral distance metric for head-related transfer function quality prediction.

Abstract

Given the substantial time and complexity involved in the perceptual evaluation of head-related transfer function (HRTF) processing, there is considerable value in adopting numerical assessment. Although many numerical methods have been introduced in recent years, monaural spectral distance metrics such as log-spectral distortion (LSD) remain widely used despite their significant limitations. In this study, listening tests were conducted to investigate the correlation between LSD and the auditory perception of HRTFs. By distorting the magnitude spectra of HRTFs across 32 spatial directions at six levels of LSD, the perceived spatial and timbral attributes of these distorted HRTFs were measured. The results revealed the limitations of LSD in adequately assessing HRTFs' perception performance. Based on the experimental results, a perceptually enhanced spectral distance metric for predicting HRTF quality has been developed, which processes HRTF data through spectral analysis, threshold discrimination, feature combination, binaural weighting, and perceptual outcome estimation. Compared to the currently available methods for assessing spectral differences of HRTFs, the proposed method exhibited superior performance in prediction error and correlation with actual perceptual results. The method holds potential for assessing the effectiveness of HRTF-related research, such as modeling and individualization.