It's because everyone is trying to use generic models for every task which is obviously terrible. If you create a custom, naroscope model, you can do some surprising things. But that takes knowledgeable employees, time, and money, none of which companies want to do. Train ann llm exclusively on recordings of drive-thru interactions and it would probably end up being quite good at it.

I mean it wouldn't hurt to also use some microphones that don't sound worse than Dollar Store Windows 98 white beige desktop microphone but that's a different conversation