This paper presents a framework to create new music using generative deep learning. Specifically, to create Western classical piano music based on music from Bach, Schubert, Chopin, and other famous composers. The core idea introduced is to use advanced Variational Autoencoders (VAE), Diffusion Models, and Generative Adversarial Networks (GAN) and combine them to create a better-performing model capable of generating classical piano music finetuned to the specific composers. The proposed ensemble methodology involves initial separate pretraining of a Variational Autoencoder (VAE) and a Generative Adversarial Network (GAN) on the dataset. Subsequently, the VAE encoder is utilized to establish a latent space, which undergoes refinement via a Diffusion Model. This refined latent space is then employed by the generator component, previously trained within the GAN framework, to generate novel music compositions adhering to the stylistic characteristics associated with specific music composers.
Unified Framework for Classical Music Generation Using GANs, VAEs, and Diffusion
2024-12-05
302279 byte
Conference paper
Electronic Resource
English