A Melodic Similarity Measure Based on Human Similarity Judgments

Music software applications often require similarity-finding methods. One instance involves performing content-based searches, where music similar to what is heard by the listener is retrieved from a database using audio or symbolic input. Another instance involves music generation tools where compositional suggestions are provided by the application based on user-provided musical choices (e.g. genre, rhythm and so on) or samples. The application would then generate new samples of music with varying degrees of musical similarity. Although several similarity algorithms such as edit distance methods and hidden Markov models already exist, they are not fully informed by human judgments. Furthermore, only a few studies have compared human similarity judgments with algorithmic judgments. In this study, we describe an empirically derived measure, from participant judgments based on multiple linear regression, for determining similarity between two melodies with a one-note change. Eight standard melodies of equal duration (eight notes) were systematically varied with respect to pitch distance, pitch direction, tonal stability, rhythmic salience, and melodic contour. Twelve comparison melodies with one-note changes were created for each standard. These comparison melodies were presented to participants in transposed and non-transposed conditions. For the non-transposed condition, predictors of similarity were pitch distance, direction and melodic contour. For the transposed condition, predictors were tonal stability and melodic contour. In a follow-up experiment, we show that our empirically derived measure of melodic similarity yielded superior performance to the Mongeau and Sankoff similarity algorithm. We intend to extend this measure to comparison melodies with multiple note changes.