Feature Selection and Stacking for Robust Discrimination of Speech, Monophonic Singing, and Polyphonic Music