Automatic Evolutionary Music Composition Based on a Reference Song

摘要

In this paper, we proposed a evolutionary music composition method based on a reference song. The proposed system automatically compose melody which is similar style with the reference. Hardwired fitness function which is independent to specific style, guarantees a quality of the compositions. Extracted features from the reference song adjust the parameters of the fitness function to compose music of specific style. Therefore, the system takes advantage both of the rule-based system and machine-learning techniques for automatically generating music. We used the k-medoid clustering algorithm with Dynamic Time Warping (DTW) distance metric to extract features from the representative melody of each cluster. From the results of the clustering, we obtained transition matrix between two clusters. We could generate a new sequence of the clusters to compose a melody, using the transition matrix. To keep consistency of the generated music, we employed copy-mutation operator between two melodies whose cluster number are same. Experimental results showed learning and generating process to compose a new melody. The composition result is presented in the musical score.