With played and recorded performances

With Borodin, we have:
a = 84.2% of played errors have been detected.
b = 73.7% of played errors have been detected and well classified.
c = 11.5% of good notes have been recognised as errors.

With Czardas, we obtain:
a = 70.3% of played errors have been detected.
b = 48.6% of played errors have been detected and well classified.
c = 9.5% of good notes have been recognised as errors.

There are two important things to notice. First (it's not a surprise), it is more difficult to do a correct classification when we have a real interpretation: Some parts of the performance are not really ``clean'' and can lead to a bad decoding of the alignment. It is also logical to observe that it is even more difficult if the piece is complex and fast.

Second point is the high value of $c$. For the same reasons, when the performance is not clean (with overlaps between notes and short added rests) we may detect errors where a human hear would not do. Especially, it happens when the follower outputs g-states when we don't have any error.

Figure 6.5: Results for real pieces with keyboard played performances