Resources for medieval musicology and liturgy
As a cross-validation for our methods to produce Mozarabic melodies, we uploaded two versions of the Roman-Frankish offertory "Scapulis suis" for the first Sunday in Lent. The first version was produced with method 1 (segmentation), the second version with method 2 (machine learning). The original manuscript runs along with the chants: the late ninth-century Laon gradual (F-LA 239), the famous contemporary of the famous Mozarabic León antiphoner (E-L 8).
When you compare these chants with the editions of Rebecca Maloy and Anton Stingl, you'll notice differences, but also apparent agreements. Among other things, the differences are due to differences in notational detail between Laon and e.g. Beneventum. We are still working on both methods.
If I understand it correctly, the neumes of the offertory 'Scapulis suis' in L-239 were subjected to the segmentation and machine learning procedures to determine the most likely melody, after which the resultant melodies were compared with the melodies according to Rebecca Maloy or Anton Stingl? Am I correct?
If my interpretation is right, the results are clearly impressive, especially with respect to the use of the segmentation procedure.
Hi Dirk, Yes you are correct. Except for the comparison with the two editions. That's for you to do.
The "impressive" results of method 1 must be seen in perspective. When the "exact" melody of Laon would have been included in our data set, method 1 simply would have found it. Since we don't know this melody, it is not included in the data set. Instead the exact Beneventan melody is included, and because it is very close to the "lost" melody, melody 1 is very close to the Beneventan one.
In fact, for me, the results of method 2 are even more impressive, since they are only based on "statistics".