Spherical-harmonics-based sound field decomposition and multichannel NMF for sound source separation

In the context of source separation solutions for virtual reality applications, several techniques in the spherical harmonics domain have been proposed in the literature. The performance of such methods is limited under high reverberation conditions and the rendering of the obtained spatial sound is fixed to the recording location only. Recently, novel sound field works in the literature proposed a global representation that enables both the direct sound (exterior field) estimation and the reconstruction in locations different from the acquisition ones. In this paper, we propose a signal processing framework based on Multichannel Non-Negative Matrix Factorization in the spherical harmonics domain that operates directly over the exterior field coefficients enabling the reconstruction of the direct sound field of the separated sources. To evaluate our proposal, we compared with other state-of-the-art source separation approaches using several setups and including different reverberation conditions, showing promising results in terms of BSS_eval metrics.