6DoF Audio Rendering Comparison
Listening experiment comparing five rendering methods for walking through a multi-point Higher-Order Ambisonics orchestra scene in VR, with open data and analysis code.
In six-degrees-of-freedom (6DoF) virtual reality a listener can walk through a captured sound scene, but a recording exists only where microphones stood. Every other position has to be rendered by interpolation or extrapolation, and the method used decides what the moving listener hears. This study compares five rendering approaches on a multi-point Higher-Order Ambisonics (HOA) capture of a virtual orchestra: a 3DoF rotation-only baseline and four 6DoF methods, based on parametric decomposition, loudness crossfading and barycentric interpolation.
The scene was synthesised from anechoic orchestral recordings with modelled directivity and room acoustics, and captured by 27 virtual third-order Ambisonics microphones, so source, room and capture geometry are known exactly. Listeners moved freely on a 6DoF head-mounted display, localised a soprano soloist (Donna Elvira’s aria from Mozart’s Don Giovanni) and rated overall quality in a MUSHRA-like protocol. 36 listeners took part at two sites, 27 at Gdańsk Tech and 9 at Zylia in Poznań; 35 are analysed for localisation and 34 for ratings.
The rendering method significantly affected both measures, but not in the same direction. The 3DoF baseline localised best, and every 6DoF method had two to four times its error. A simple barycentric interpolation, with three times the baseline’s localisation error, was rated no differently from the baseline. The multi-receiver parametric renderer, run without its object stage to keep real-time processing stable, did worst on both measures. Localisation accuracy and perceived quality can therefore come apart in navigable VR audio, and the paper derives design guidance by application priority.
The repository holds the raw experiment logs, unmodified, and an R Markdown analysis that generates every statistic, figure and LaTeX value in the paper: a Gamma mixed model for localisation error, a linear mixed model for the ratings, estimated marginal means with post-hoc comparisons, and simulation-based residual diagnostics. It also reports sensitivity analyses with site as a fixed effect, without the second site, and without the author’s own session. Data and code are released under CC BY 4.0, and the v1.0.0 release on Zenodo is the version used for the paper.
Sole author. Submitted to IEEE Transactions on Visualization and Computer Graphics on 8 November 2026, under review.