Finding occurrences of melodic segments in folk songs employing symbolic similarity measures

Much research has been devoted to the classification of folk songs, revealing that variants are recognised based on salient melodic segments, such as phrases and motifs, while other musical material in a melody might vary considerably. In order to judge similarity of melodies on the level of melodic segments, a successful similarity measure is needed which will allow finding occurrences of melodic segments in folk songs reliably. The present study compares several such similarity measures from different music research domains: correlation distance, city block distance, Euclidean distance, local alignment, wavelet transform and structure induction. We evaluate the measures against annotations of phrase occurrences in a corpus of Dutch folk songs, observing whether the measures detect annotated occurrences at the correct positions. Moreover, we investigate the influence of music representation on the success of the various measures, and analyse the robustness of the most successful measures over subsets of the data. Our results reveal that structure induction is a promising approach, but that local alignment and city block distance perform even better when applied to adjusted music representations. These three methods can be combined to find occurrences with increased precision