Multimodal input is not as simple as "stacking on features"; when the primary channel fails, the backup channel must be robust enough, otherwise it is merely failing in a different way.