I'm developing a scanned document recognition system from long ago and I still have to improve the use of Tesseract to make it usable. I have done a lot preprocessing but am not able to improve the accuracy.
I have some experience with tesseract. In my case doing some morphology to merge disjoint parts of letters helped (We used bank tickets printed with matrix printers).
In addition, and I think this is the most important. We trained for the correct typography, and that boosted from 50% to 90% in the performance.
My pleasure I hope it helps. The training process is tiresome, but works. Be sure to have good samples of every letter for each typography present in your documents.