COMBINING MULTILEVEL AND MULTIFEATURE REPRESENTATION TO COMPUTE MELODIC SIMILARITY

In the proposed approach, melodic similarity is computed as a content-based information retrieval task. To this end, the initial incipit is considered as the query in a query-byexample paradigm and the ranked list of potentially similar documents is given by the list of retrieved documents. The approach to retrieval is based on document indexing, where each document is described by alternative melodic features of note substring with different lengths. The document collection is separetely indexed for each feature. The similarity between the query and the other documents is computed only on indexes, obtaining a number of rank lists of potentially similar documents. The final result is given by the application of a data fusion technique on the single rank lists..