Separation of direct sounds from early reflections using the Entropy Rate Bound Minimization algorithm

Blind Source Separation (BSS) finds applications in audio scene analysis to identify and separate sources from instantaneous or convolutive mixtures. Independent Component Analysis (ICA) is notably a powerful tool for BSS allowing to decompose a multichannel recording into a set of independent components. However, real acoustic recordings also contain sound reflections which can be considered as secondary sources highly correlated to the direct sounds. In this paper, we propose to study the ability of the Entropy Rate Bound Minimization (ERBM) algorithm to separate direct sounds from reflections. The evaluation is conducted through a simplified audioconference scenario simulating a 2nd-order ambisonic microphone sound capture. Results are compared to classical ICA algorithms, including tensorial methods or entropy minimization algorithms. Objective measures show that, thanks to some assumptions on the sources' model, the ERBM algorithm outperforms state-of-the-art methods, hence showing ability to separate cross-correlated speech sources and then direct and reflected signals.