Strange that they don't normalize both the input and reference texts, I do a lot of work with text from different sources and I mostly work with normalized versions of them.