Why would they need to use ML to find intros? Most implementations I have heard of just analyze the sound at the beginning of episodes and mark out the "theme song" that recurs.
Hmmm, well it might be that my memory is faulty and/or I read someone else's comment somewhere when the feature first launched that said it was using ML. But you are correct, both the initial blog [0] and the support article [1] makes no mention of ML, only audio fingerprinting. My mistake.
