I have a dataset with short melodies hummed by experimental subjects. Each melody consists of a variable number of notes (10 to 20 notes) and is coded by the pitch levels of the notes (automatically extracted from the audio file), and also in terms of its pitch contour (as Parsons code, e.g. "*udduududuudu").
For each pair or melodies, I would like to compute a numerical measure of contour similarity, which amounts (I think) to finding a suitable measure of string similarity between the corresponding Parson codes.
I initially used a simple edit distance (specifically, this implementation), however I am not sure this is "robust" enough, given the noisy nature of the automatically extracted pitch levels of the melodies.
Can anyone advise if the edit distance (the implemention I used of it) is suitable as a measure of string similarity, or if another measure might be more suitable?