I'm sure they did. And probably still do. But no amount of paid for training data is going to cover all situations that occur in real life. Accents, mannerism, speech imperments - you can't cover all possible premutations.