The speed issues for similar_text seem to be only an issue for long sections of text (>20000 chars).
I found a huge performance improvement in my application by just testing if the string to be tested was less than 20000 chars before calling similar_text.
20000+ took 3-5 secs to process, anything else (10000 and below) took a fraction of a second.
Fortunately for me, there was only a handful of instances with >20000 chars which I couldn't get a comparison % for.