Deep Within-Class Covariance Analysis for Robust Audio Representation Learning