Comparing Ambisonics-based Binaural Rendering Methods: Can Perceptual Models be of Use?
* Presenting author
Abstract:
The Ambisonics framework is often used to produce binaural audio from microphone array recordings or real-time audio engines. Typically, Ambisonics signals are processed at low spatial orders due to hardware constraints or as a way to reduce complexity, which generally results in a degradation of the quality of the binaural output. There exist several methods that alleviate this issue but they have not yet been compared for a wide range of spatial orders and perceptual metrics, because that would involve very extensive and time-consuming listening experiments. We investigated to what extent binaural models can be used to perform such an evaluation in a rapid and consistent manner. Several state-of-the-art binaural Ambisonics rendering methods were compared using binaural models for spatial orders 1 to 44 in terms of localisation performance, externalisation and speech reception in noise. The models predicted an overall increase in performance with spatial order, as expected. A clear effect of the rendering method was observed: at high orders, most methods converged to a good performance, whereas at low orders, some methods produced significantly more accurate binaural signals than others. These predictions, supported by a numerical analysis, have been found to be in line with results from previous literature.