DrumGAN - Adversarial synthesis of drum sounds

Abstract: This paper faces the problem of audio synthesis with Generative Adversarial Networks (GAN), with an attempt to create novel, original high-quality samples of drums that could be used within the realm of music production. My results show that it is possible to create high-quality drum samples with architecture such as DCGAN (Deep Convolutional GAN) and that using causal and dilated convolutions is a viable approach, while not making clear if this approach is significantly better than the one of using standard convolutions. It also shows that 1-d transpose convolutions can be substituted with nearest neighbour upsampling followed by regular 1-d convolutions for GANs that generate one-dimensional data. In the final discussion the idea of initializing generator weights in a strategic way in order to increase GAN training stability is introduced.