I think it's been mostly throwing hardware at the problem. Server hardware, backed by terabytes of context data. I remember back in the day, most of the errors

I think it's been mostly throwing hardware at the problem. Server hardware, backed by terabytes of context data. I remember back in the day, most of the errors were the exact same errors that a human would make. Even you and I only really hear 95% or so of the words someone says. But we can fill in the rest from the context. It always amazes me when I dictate some sentence to my phone, and at the end I see one of the words change to another that sounds almost identical but makes much more sense in the entire context of the sentence.