A computational approach to content-based retrieval of folk song melodies

In order to develop a Music Information Retrieval system for folksong melodies, one needs to design an adequate computational model of melodic similarity, which is the subject of this Ph.D. thesis. Since understanding of both the properties of the melodies and computational methods is necessary, this problem requires a multidisciplinary approach. Chapter 2 reviews the relevant academic background of both Folk Song Research (as sub-discipline of Ethnomusicology) and Music Information Retrieval. It also presents an interdisciplinary collaboration model in which Computational Musicology serves a ‘man-in-the-middle’ role, with the particular task to design computational models of concepts from Musicology, in this dissertation especially the concept of tune family. An important step towards the understanding of the concept of tune family is the method to annotate similarity relations between melodies that is presented in Chapter 3. Its aim is to make aspects of experts’ intuitive similarity assessments explicit. 360 melodies in 26 tune families were ‘manually’ annotated, resulting in an Annotated Corpus that is a valuable resource for the study of melodic similarity and for the evaluation of computational models of melodic similarity. From the annotations we conclude that the relative importance of the various dimensions of melody varies to a large extent in individual comparisons. Furthermore, it appears that in many cases melodies are judged to be related based on shared characteristic melodic motifs. In Chapter 4, 88 low-level, global, quantitative features of melody are used to discriminate between tune families. It appears that such features can be used to recognize melodies within the relatively small Annotated Corpus, but that they lose their discriminative power in a larger dataset of thousands of melodies. Chapter 5 uses the same kind of features to assess authorship problems of fugues that are in the catalogue of J.S. Bach. Hypotheses from musicological literature could be supported. The various degrees of success of the same computational method in the previous and the current chapters show that computational methods cannot blindly be applied to musicological questions. In Chapter 6, the potential of alignment algorithms for folk song melody retrieval is studied by incorporating musical knowledge in the algorithm in the form of appropriate, musically motivated, substitution scoring functions. This approach leads to good retrieval results both for a small (360 melodies) and a large (4830 melodies) dataset. Furthermore, domain experts were able to classify ‘problematic’ melodies using the results of alignment algorithms. This thesis contributes both to Folk Song Research and Music Information Retrieval by incorporating musical knowledge in computational models. The process of developing such models leads to better understanding of melodic similarity and, thus, of the concept of tune family, which is relevant for Folk Song Research. The models that have been developed, have successfully been used for melody retrieval. From the research that is presented in this thesis, it is clear that computational methods have a rich potential for the study of music; not as a replacement of ‘traditional’ methods, but as an extension of the research methods that are available to the musicologist