I'm really not sure a CRC would be good : https://stackoverflow.com/questions/3645461/creating-a-fast-hash-function-for-fixed-length-inputIt is but it's unnecessarily complex for this problem. We can simplify the problem because the input "text" is constant length. In fact I'd be willing to wager a 32-bit CRC would provide sufficient uniqueness.
Computing md5 is very fast and quite good
Anyway i guess you can try the CRC32 method on a very huge IR collection and see if you get collisions ?