Late Fusion Ensembles for Speech Recognition on Diverse Input Audio Representations