Investigating the perceptual accuracy of machine-learning generated personalized head-eelated impulse responses

Abstract

An accurate pair of head-related impulse responses (HRIR) is necessary to render a realistic and immersive virtual auditory space. Since personalized HRIRs are challenging to measure, a method of using two machine-learning models to separately estimate the shape and time delays of the HRIRs based on an individual's anthropometric measurements of the head and pinnae is proposed. The investigation evaluates the perceptual accuracy of the resulting HRIRs with two subjective listening studies. First, an ABX test was conducted to determine if listeners could distinguish between auralizations made with an actual HRIR from the CIPIC database and the personalized HRIR based on the corresponding anthropometric measurements. In the second study, subjects performed localization tests in virtual reality to determine the perceptual accuracy of their personalized HRIR compared to an average HRIR. Each set of tests included one of three auralizations (noise, music, or speech) played in five source locations. Perceptual accuracy was evaluated by comparing the localization errors for the average and estimated HRIRs. The stimuli were repeated four times to investigate if subjects improved over time. The subjects were also asked to rate their listening experience to assess factors such as externalization and difficulty of localization.