Melody generation system based on a theory of melody sequence

We propose a melody generation system based on the Implication-Realization Model (IRM) of music theory. The IRM is a music theory, which was proposed by Eugene Narmour. The IRM abstracts music. It then expresses music according to symbol sequences based on information constituting the music pitch, rhythm, and rests. Previous melody generation systems are mostly based on tone transition models, which do not have function of abstracting melodies observed in training data. In those previous systems, generated melodies do not reflect tone sequences that do not exist in training data. However, it is obviously required that a melody generation system is able to abstract melodies in training data and to output certain melody, which is rarely observed in the training data. Our melody generation approach properly abstracts melodies in training data based on the IRM. The IRM expresses contexts of melodies using symbol sequences. Our melody generation system consists of two models; that of symbol sequence transition and that of generating tones from symbols. With the former model, the symbol transition probability model is trained with the results of the IRM analysis. The system then generates an optimal symbol sequence according to the probability model. Then, from a set of tones, each symbol sequence generates a melody. We evaluated the proposed system through subjective human judgment and the results showed that our system properly generated melodies.