Melody2Vec: Distributed Representations of Melodic Phrases based on Melody Segmentation

Abstract

In this paper, we present melody2vec, an extension of the word2vec framework to melodies. To apply the word2vec framework to a melody, a definition of a word within a melody is required. We assume phrases within melodies to be words and acquire these words via melody segmentation applying rules for grouping musical notes called Grouping Preference Rules (GPR) in the Generative Theory of Tonal Music (GTTM). We employed a skip-gram representation to train our model using 10, 853 melody tracks extracted from MIDI files primarily constructed from pop music. Experimental results show the effectiveness of our model in representing the semantic relatedness between melodic phrases. In addition, we propose a method to edit melodies by replacing melodic phrases within a musical piece based on the similarity of the phrase vectors. The naturalness of the resulting melody was evaluated via a user study and most participants who did not know the musical piece could not point out where the melody had been replaced.