Unlike most spell checking systems that just look up known words in a list, Paratext uses algorithms to break down the existing text and generate a “language profile”. This language profile is then used to determine how likely it is that a word is misspelled. Above a certain threshold, it will display it as an error. It then learns based on how people correct words throughout the project’s lifetime to get smarter about spelling. This complicated system is needed because we work with vernacular data where lists of correct-words don’t yet exist.
The advantage of this system is that it can start doing fairly decent spell-checking with very little text and it will get better as work on the project continues. The disadvantage is that it can take a long time for the algorithms to break down the workings of the project’s language to create the language profile.
That being said, the language profile information is cached, so after the first time it’s run, it should be much faster the next time. Also, if I recall correctly, the spell checking did have a large performance improvement in 7.6 for certain circumstances. If the performance makes it unusable in 7.5, you might try upgrading to 7.6 or 8.0.
Note that, although the language profile information is cached, the language profile will be rebuilt once a month (I think) to make sure it’s updated with any new information from any new text that has been added to the project.