The efficiency comes from the lack of voice processing. The beeps and boops are easier on CPU resources than trying to parse spoken word.

That said, they should just communicate over an API like you said.