Letter recognition can only be done in conjunction with a spelling application. In other words, if you have a list of street names, spelling
W O O D W A R D
will work, if you use an n-best list and search through the dictionary for the results. As long as your task can be constrained by a dictionary, even a huge dictionary, you’re OK. You’ll have to patch OpenEars for N-best output, though, and build an FSG or LM. The LM will work better if you build it from your dictionary.
If your letter sequences truly are random, you’re dealing with something that’s beyond the state of the art. It’s beyond the state of the art for human listeners, too. Give it a try, read some random letter sequences to people and see how many they get wrong.
Is this something you’re still working on?