Recent advancements in deep learning technology have made significant strides in the field of musical creation based on GAN architectures. GAN algorithms consist of two parts: generator and discriminator, where through their mutual competitive training process, the generator develops techniques in creating musical pieces closer to the properties of the real musical pieces. In this project, a classical multi-track music generation architecture named MuseGAN will be used to increase the context similarity of generated musical phrases through modifications in generator’s temporal architecture. Additionally, the use of a feature extractor, along with other modifications during the training process, will improve the smoothness of the transition between notes. The evaluation of the MuseGAN generated music samples uses several parameters: distribution density of notes, chord matching score, harmony, and BLEU score as well as Self-BLEU. The results prove that the generated music samples using the MuseGAN algorithm show high harmony score and good musicality. The mutual interaction between generator and discriminator improves music created by artificial intelligence, making it more diverse and realistic.
