You could try using Text::LevenshteinXS which on previous benchmarking is considerably quicker than the pure perl version.

However, using character-based comparison measures against text made up of words and sentences is never going to be particularly effective.