Another option would be to use a parametric speaker array for focused, highly-directional audio. Then, rather than aim it at an individual person, bounce the audio off of the nearest suitable surface (a window, perhaps). They'd never know where the voice was coming from, and might even turn on each other in their confusion.