Sequence-to-Sequence Models Can Directly Translate Foreign Speech

   Abstract