90% of radio communications is standard and expected.

If I get the call sign and a few key words, I can normally figure it out. For the most part you kinda know what people are gonna say before they say it, or at least have it narrowed down to just a few possibilities. You also just get an ear for the narrowband AM sound after a few hours.

“(Garbled) n123a (garbled) joining (garbled) one three” heard at an untowered field is going to be “local traffic, Cessna N123a is joining the downwind leg for runway 13.”

Or if you’re on final approach and the tower calls you up, there are really only a few things they are going to tell you, most likely is “Cessna 123a cleared to land”.

Plus, I’m guessing you have o my listened to very busy feeds at commercial airports. Students at smaller airports will often announce that they are a student, and controllers and other pilots will try to slow it down and give them a little extra room and time.

No idea how AI does it, but I assume it’s not too hard since there are essentially unlimited hours of recording to train on.

i'm guessing the model is doing something similar, and has a fairly confident set of the possible phrases to map sounds to.