Speaker
Description
The 2024 Listener Acoustic Personalisation (LAP24) challenge benchmarked the spatial upsampling (interpolation) of head-related transfer functions (HRTFs). Seven teams submitted HRTFs that were upsampled from four sparse grids containing between three and one hundred HRTF positions. The submissions included six deep learning-based approaches and one algorithmic approach. They were evaluated with respect to log-spectral distortion (LSD) and broadband interaural time and level differences (ITD, ILD). The learning-based upsampling methods often showed smaller differences to the reference HRTF than the algorithmic approach, especially for very sparse sampling grids. In the current study, we conducted a complementary perceptual evaluation of the LAP24 challenge algorithms with respect to colouration and source direction for a static sound source. This work confirmed the numerical results of the LAP Challenge at least for the top-ranked method. Beyond first place, however, the perceptual ranking differed from that obtained with physical error measures, due to the different aspects captured and their sensitivities. Moreover, an algorithm-specific diffuse field equalisation filter had little effect, suggesting that direction-dependent upsampling errors remain audible even in this case. The upsampled HRTFs, along with audio examples and ratings from the listening test, are publicly available to facilitate benchmarking of future upsampling algorithms.