State-of-the-art Speech Recognition With Sequence-to-Sequence Models