>I suspect an LLM that supports audio input would outperform most drive through window attendants.

I think this would lead to a lot of really frustrating errors in a food ordering context. In my experience LLM speech recognition is really prone to inserting or removing negation words, which can really shift the meaning. This is a problem in any speech recog use case but there's often sufficient context to understand when the error has occurred. In a list of what a customer does / doesn't want, there is no additional context to fill the gap.

"Cheeseburger with ketchup mustard onion no tomato" -> "Cheeseburger with ketchup mustard onion tomato" or the inversion

You can correct this to some degree with an order confirmation step but the error rate would mean a lot of customers need to make a correction, which adds time to the process and increases customer frustration.

At that point I'd rather the drive thru have a touchscreen and just let me input the order directly, versus having to argue with an LLM about which ingredients i want included or excluded.