부산대학교 도서관

The generation of music by neural network is an interesting research topic. Since songs are composed of long-term structures of melody, rhythm, and chord progression, the construction of a good music generation model is challenging. This paper proposes a music generation scheme that uses aligned beats to fuse two segments of music. First, a certain segment in the song is selected as the starting segment of the song. In order to take into account the relationship between notes and the relationship between the whole bar, the music generated by two music generation schemes is used. The two different music segments generated by the two schemes are used to learn the fusing weight of each beat so as to generate new fused melodies with coherence and harmony. The experimental result generated by the model are evaluated using an objective evaluation mechanism. The performance results illustrate that, the proposed scheme outperforms the other music generation schemes in terms of the Pitch Class Consistency (PCC), Note Length Consistency (NLC), Grooving Consistency (GC), and Limited Macroharmony (LM).