It's an established fact that Apple has traditionally refused to just use all of their user recordings as training data, and is willing to accept worse voice recognition as a result.