The difficult part is training a model like this - it's very expensive to train a huge model on video, you need lots of work and infrastructure, and you can't legally do any of this in the first place.
You're human being capable of understanding what you're seeing, ML is just computational statistics. You need to throw a ton of data at it to get accurate results.
You might be able to piggyback on pre-trained models via transfer learning in order to reduce the amount of data and resources needed to train a model for this task, though.
I don't think Shazam uses ML for song recognition, but instead hashing and fingerprinting via Fourier transforms.