A Robust Subspace Classification Method for Highly Correlated Acoustic Signals

This paper proposes a subspace based classifier, which can separate highly correlated acoustic signals based on source material. In this method, the optimum set of Eigen-filters that form the subspace classifier are selected such that the cross correlation between different classes is minimized. The proposed method has high noise immunity as the noise subspace is eliminated at the subspace separation stage. Then the resolution of the subspace classifier is varied and its impact is analyzed for the given set of signals. Finally, robustness and the practicality of the proposed classifier is verified by applying it for two application scenarios, namely, "decision making in cricket" and "hidden information extraction from speech signals in order to reveal the speaker identity".