It wouldn't be too hard to train. There are enough audio models and computer vision models that could be trained in parallel on video clips that have recorded sound to train what sound profiles are associated with what events in the frame.

The real fun one would be to figure out how to train an AI to understand sounds originating from out of frame.

It won't be long before we have something like Oobabooga or Stable Diffusion but for artificial video with matching audio. I'm so sorry for historians in the future trying to determine whether this video of Joe Biden doing tap dances on an F-16 while throwing a hadouken recovered from a trashed hard drive is authentic or not.