Search

The aim of this paper was to provide a derivation to explain the performance of the second-order-based blind signal separation (BSS) in reverberant environments. In particular, the second-order-based BSS algorithm, which exploits non-stationarity of the input signals, is investigated. The derivation provides a quantitative link between the reverberation parameter of the environment with the cost function of the BSS. Importantly, the theoretical finding complements existing literature on the interpretation of BSS and provides incremental insight to its performance. [ABSTRACT FROM AUTHOR]

Blind source separation (BSS) method separates the desired signals from a mixed observed signal by making full use of spatial information. Spatial information refers the fact that the sources originate from different location in space and thus provides the diversity for the separation. Similarly, this diversity can be further enhanced by optimizing the sensor placement as opposed to a fixed location. This paper aims to fill this research gap by proposing a sensor placement optimization strategy to further improve the performance of BSS. As the problem is non-convex in nature, a new hybrid descent optimization method is proposed by embedding a gradient-based method into the genetic algorithm. The proposed method benefits from the robustness of the genetic algorithm and the fast convergence speed of the gradient-based method. Results show that the optimized sensor placement greatly improves the separation performance of the BSS system across the different reverberation times. [ABSTRACT FROM AUTHOR]