Just out of interest, what model are you running?

Are you giving the model context? Industries XYZ with special words zyx

In our use case it's better then the human transcript comparison