In testfiles where we only have note errors, results are:
a = 93.5% of played errors have been detected.
b = 74.2% of played errors have been detected and well classified.
c = 2.5% of good notes have been recognised as errors.
Detection is not a problem, but classification remains uncertain, especially in case of consecutive errors. In our case, we have some good results with isolated errors, but if we have more than 3 consecutive wrong notes we obtain a bad classification.
This problem is linked to the size of the used piece: making more than 3 errors on a 8 notes piece does not really allow to have a good alignment...