What can machine translation learn from speech recognition?