Geometrically Constrained Joint Moving Source Extraction and Dereverberation Based on Constant Separating Vector Mixing Model

This paper proposes a joint optimization algorithm that simultaneously achieves geometrically constrained moving source extraction and dereverberation. Recently, Constant Separating Vector (CSV) mixing model has been proposed, which is useful for moving source extraction. Based on this model, Independent Vector Extraction with auxiliary function (CSV-AuxIVE) has been proposed as a fast and stable moving source extraction method. When applying CSV-AuxIVE to a more realistic situation, we need to avoid the uncertainty of which signal CSV-AuxIVE extracts. Moreover, its source extraction performance will be limited when the reverberation time is long. Thus, to extract the moving target source signal in a highly reverberant environment, we derive an algorithm that jointly optimizes Weighted Prediction Error (WPE)-based dereverberation and CSV-AuxIVE with Geometric Constraint (GC), which we call GC-CSV-WPEIVE. By applying several GCs into the range of the moving target speaker, our proposed algorithm can extract the target signal, even in a highly reverberant environment. We show the efficacy of GC-CSV-WPEIVE using a simulation experiment. Furthermore, we discuss the robustness of its extraction performance against the moving range estimation error.