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.