Blind source separation of nondisjoint sources in the time-frequency domain with model-based determination of source contribution

Abstract

While most blind source separation (BSS) algorithms rely on the assumption that at most one source is dominant at each time-frequency (TF) point, recently, two BSS approaches, [1], [2], have been proposed that allow multiple active sources at time- frequency (TF) points under certain assumptions. In both algorithms, the active sources in every single TF point are found by an exhaustive search through an optimization procedure which is computationally expensive. In this work, we address this limitation and avoid the exhaustive search by determining the source contribution in every TF point. The source contributions are expressed by a set of posterior probabilities. Hereby, we propose a model-based blind source separation algorithm that allows sources to be nondisjoint in the TF domain while being computationally more tractable. The proposed BSS approach is shown to be robust with respect to different reverberation times and microphone spacings.