Current behaviour

When an estimated key is not the same as an annotated key, there are different ways of evaluating how "far off" the estimated key is.

Currently, in madmom.evaluation.key.error_type the categories outlined in the MIREX competition are used: 'fifth', 'relative', 'parallel' and 'other'.

Also, directly associated to the type of error, a score is given based on the MIREX key estimation guidelines.

Desired behaviour

1. It would be nice to enable experimenting with other sets of scores for the different error categories; right now they are strongly coupled in the code and it would be nice to dissociate those.
2. It would be nice to separate key detection errors where the detected key is the relative key of the fifth of the annotated key. For example, if the annotated key is a C major and the estimated key is an E minor, there is only one note that is different between the two scales (F#), just as in G major (which is the fifth of C major). It would then seem logical to provide its own error type to that case as it is still a closely related key compared to the annotated key.

I have submitted a PR related to that issue.