Wed-2-10-3 Discriminative Singular Spectrum Analysis for bioacoustic classification

Abstract: Classifying bioacoustic signals is a fundamental task for ecological monitoring. However, this task includes several challenges, such as nonuniform signal length, environmental noise, and scarce training data. To tackle these challenges, we present a discriminative mechanism to classify bioacoustic signals, which does not require a large amount of training data and handles nonuniform signal length. The proposed method relies on transforming the input signals into subspaces generated by the singular spectrum analysis (SSA). Then, the difference between the subspaces is used as a discriminative space, providing discriminative features. This formulation allows a segmentation-free approach to represent and classify bioacoustic signals, as well as a highly compact descriptor inherited from the SSA. We validate the proposed method using challenging datasets containing a variety of bioacoustic signals.