It's hard to get the recognition as good as I'd like, but I still have some ideas for how to improve it. As you say it's too sensitive in some cases (long curly strokes) and not strict enough in other cases. For the straight lines I guess I'd need to take into account the relative positions compared to earlier strokes and where they cross each other, which is the next thing I'll try.
I've also thought about letting the user rate themselves. But personally I'm not so fond of it, since I find myself "cheating" sometimes, and also have to put some thought into rating my performance instead of just focusing on learning. But I'll keep it in mind, maybe having some kind of mixed solution is best, where the user can override the results.