Learning Compact Structural Representations for Audio Events Using Regressor Banks