Monochord Melodic Similarity Retrieval Evaluation and Applications for Composer Classification

Abstract

The RISM A/II database is filled with the musical notations of the beginnings of more than a million melodies. The Monochord search engine can retrieve melodies that are similar to a query melody using several search methods, amongst which pitch raters, weight-based raters and duration-based raters. The performance of all 27 search methods is evaluated using mean average precision metrics and the TREC framework that is suited for retrieval performance analysis. The difference in exact pitch between melodies turns out to be the best factor to search with for musical similarity retrieval. All melodies have metadata such as a composer name, but a portion of the database is labelled as Anonymus. A k-Nearest Neighbours algorithm is optimised for the purpose of deanonymisation and used to classify several Anonymus songs to test the applicability of this classifier for composer labelling. Using a classifier for deanonymisation purposes turns out to be viable with human correction.