Listener Acoustic Personalization Challenge—LAP24: Head-Related Transfer Function Dataset Harmonization

Big data analysis and collation for data-driven head-related transfer function (HRTF) personalization methods are often hindered by systematic differences between HRTF datasets. To address this issue, we designed Task 1 of the inaugural listener acoustic personalisation (LAP) challenge. Researchers were invited to propose strategies for harmonizing HRTFs from a collection of eight different datasets so that dataset-specific artifacts were mitigated while preserving the perceptually relevant attributes of the original HRTFs. Defining the two-sided task required a deeper assessment of the acoustic and perceptual HRTF descriptions to find an evaluation framework that encompassed the two domains. Consequently, a two-stage evaluation was devised to assess the submissions. First, an auditory sound localization model was used to test the perceptual validity of the harmonized HRTFs by estimating the difference in sound localization performance between the original and the harmonized versions. Then, a machine learning classifier was employed to differentiate harmonized HRTF datasets, and its accuracy was used to rank submissions. Three submissions were received, and one was declared a winner according to the evaluation criteria. Further analysis of the submissions revealed some limitations of the evaluation system, prompting a comprehensive review of the task’s inherent complexities. This paper serves as a systematic account of the challenge and relevant considerations, intended to guide future advancements in the field of HRTF personalization research.