MidiPGAN: A Progressive GAN Approach to MIDI Generation

While recent research in music generation has mostly focused on encoder decoder architectures and self-attention mechanisms, prominent advancements regarding GANs have not yet been incorporated for the creation of music. These include solutions for major challenges when training GANs, most importantly training instability. In this work, we aim to apply this new knowledge to music generation, in order to make it more efficient and enable the automatic creation of music of higher quality. We utilize the progressive approach towards GANs, and implement it to train on symbolic music data. For best results, we process this data to obtain a new dataset, which matches the progressive approach. To achieve this, we propose a new way of downsampling fit for musical data. We furthermore conduct a user study to evaluate our results, and compute an FID score of 12.30 as objective metric.