A Comparison of Statistical and Rule-Based Models for Style-Specific Harmonization

The process of generating chords for harmonizing a melody with the goal of mimicking an artist’s style is investigated in this paper. We compared and tested three different approaches, including a rule-based model, a statistical model, and a hybrid system of the two, for such tasks. Experiments were conducted using songs from seven stylistically identifiable pop/rock bands, and the chords generated by the systems were compared to the ones in the artists’ original work. Evaluations were performed on multiple aspects, including calculating the average percentage of chords that were the same and those that were related, studying the manner in which the size of the training set affects the output harmonization, and examining a system’s behaviors in terms of the ability of generating unseen chords and the number of unique chords produced per song. We observed that the rule-based system performs comparably well while the result of the system with learning capability varies as the training set grows.