Optimizations of the Spatial Decomposition Method for Binaural Reproduction
PubDate: January 1, 2021
Teams: TH Koln – University of Applied Sciences；Facebook Reality Labs Research
Writers: Sebastià V. Amengual Garí, Johannes M. Arend, Paul T. Calamia, Philip W. Robinson
The spatial decomposition method (SDM) can be used to parameterize and reproduce a sound field based on measured multichannel room impulse responses (RIRs). In this paper we propose optimizations of SDM to address the following questions and issues that have recently emerged in the development of the method: (a) accuracy in direction-of-arrival (DOA) estimation with open microphone arrays utilizing time differences of arrival as well as with B-format arrays using pseudo-intensity vectors; (b) optimal array size and temporal processing window size for broadband DOA estimation based on open microphone arrays; (c) spatial and spectral distortion of single events caused by unstable DOA estimation; and (d) spectral whitening of late reverberation as a consequence of rapidly varying DOA estimates. Through simulations we analyze DOA estimation accuracy (a) and explore processing parameters (b) in search of optimal settings. To overcome the unnatural DOA spread (c), we introduce spatial quantization of the DOA as a post-processing step at the expense of spatial distortion for successive reflections. To address the spectral whitening (d), we propose an equalization approach specifically designed for rendering SDM data directly to binaural signals with a spatially dense HRTF dataset. Finally, through perceptual experiments, we evaluate the proposed equalization and investigate the consequences of quantizing the spatial information of SDM auralizations by directly comparing binaural renderings with real loudspeakers. The proposed improvements for binaural rendering are released in an open source repository at https://www.github.com/facebookresearch/BinauralSDM.