Application of a Model Based on a Variational Autoencoder for Music Generation

This study introduces a novel approach to music generation using a Variational Autoencoder (VAE) model, which incorporates style embeddings for enhanced control over the generated music. The model divides the latent space into content and style components, allowing users to specify desired musical styles. Experimentation demonstrates superior performance compared to traditional methods, with the model effectively capturing stylistic nuances and producing diverse musical compositions. Methodologically, the VAE employs a reparameterization trick and μ-forcing technique to ensure effective training and preservation of latent variables. The study concludes that the proposed approach surpasses baseline methods, offering users greater control and flexibility in generating music tailored to specific styles, thereby advancing the field of AI-driven music composition. The aim of the work is to develop a model for genre-specific music generation based on VAE neural networks.