Optimally Encoding Inductive Biases into the Transformer Improves End-to-End Speech Translation.