A novel method for data segmentation to covariance stationary regions

This paper presents a new method for segmenting non-stationary data to covariance-stationary regions that is of importance in some applications such as subspace-based speech enhancement. The proposed method utilizes a test statistic previously suggested for Synthetic Aperture Radar (SAR) image segmentation. In this paper, employing Random Matrix Theory, we derive two first moments of the test statistic which are used to define an explicit decision threshold rather than heuristic one. We show through Monte Carlo simulation some interesting properties of the test statistic. The reasonable performance of our approach is validated by means of ROC curve obtained as the result of applying the proposed method to the synthetic data. Moreover, we compare our algorithm with the two recently introduced algorithms of which one aimed at locating variance-stationary regions and the other is designed to find the data intervals with the same covariance structure. The superior performance of our algorithm as well as its low computational cost is shown in this paper.