Abstract
Head-related transfer functions (HRTFs) often play a crucial role in spatial hearing, immersive audio applications for virtual reality (VR) and augmented reality (AR), and help in improving hearing assistive devices. The Listener Acoustic Personalisation (LAP) challenge 2024 aimed at advancing research in spatial audio personalisation, with a focus on head-related transfer functions (HRTFs). The challenge was split into two tasks: Task 1 was on HRTF harmonisation, and Task 2 dealt with spatial HRTF upsampling. This paper presents the results and reports the findings related to Task 2 of the LAP challenge. The submissions to Task 2 employed both algorithmic and machine learning-based approaches, which were evaluated on three key spatial audio objective metrics, including the log-spectral distortion (LSD), interaural time difference (ITD), and the interaural level difference (ILD). The results highlighted the strengths and limitations of various upsampling techniques, with learning-based methods demonstrating superior performance at lower sparsity levels. In terms of the LSD, seven of the submissions achieved an impressive performance of less than 5 dB when upsampling from only three measurement points. The results also highlighted that most submissions were often not able to outperform a generic HRTF created by averaging the HRTFs in the training dataset. One of the main contributions of this paper is that it showcases the limitations of objective metrics when it comes to evaluating HRTF upsampling. Therefore, this paper argues that a more holistic approach is needed going forward, which should include the integration of multiple perceptually relevant measures, as this is the only way to ensure a well-rounded assessment of HRTF upsampling quality.