Non-Stationary Signals Separation Using STFT and Affinity Propagation Clustering Algorithm

In this paper, we address the problem of separating N unknown non-stationary signals using as many observed mixtures. Using short-term Fourier Transform (STFT) of the mixtures along with a classification approach based on affinity propagation (AP) clustering provide an efficient technique for separating non-stationary signals. The proposed method is featured by its simplicity and improved classification compared to other existing TF based signal separation methods. The method can tackle both the mono-component as well as multi-component signals and its assumptions about the mixing matrix are more relaxed than other existing methods. To the best of our knowledge, this is the first signal separation approach based on AP clustering. Besides improved clustering the AP does not require apriori knowledge of the number of clusters. Examples, using synthetic as well as real-life data, are presented to demonstrate the validity and efficiency of the proposed method.