In this paper, MIDI data of Korean traditional music and music from the Central Plains were collected, melodic features and note features of the music data were extracted, and a multi-style melody and arrangement generation (MSMICA) model was constructed. The Multi-Sequence Generative Adversarial Network serves as the main framework of the model, including two core parts, the generator and the discriminator. The generator incorporates a reinforcement learning mechanism to reward with feedback from the discriminator, while the discriminator utilizes the GRU component for the recognition of mesogenic musical elements as well as harmonic music generation, and proves the feasibility of the model based on relevant experiments. During the training process, the Loss and Accuracy values of the model converge to about 0.44 and 0.95 successively in about 60 rounds of iterations, showing good training results. The fluctuation amplitude of the signal features of the model-generated traditional music of the Central Plains style Korean music is between ±0.8, and the signal features in the Meier spectrum, spectral roll-off and chromatic frequency are similar to the real music, and the model-generated music clip is more in line with the traditional music of the Central Plains style Korean music.
