Music Generation Using Deep Learning Techniques

Abstract

This project aims to generate songs using the Jukebox model and its architecture. Jukebox’s Vector Quantized Variational AutoEncoder (VQ-VAE) architecture is state-of-the-art deep generative model used for music generation and gives an outstanding result. For this purpose, different Elvis Presley songs were analyzed in audio domain using various Music Information Retrieval (MIR) methods. The top level of the Jukebox model was retrained with these songs in order to increase the quality of the songs that will be produced in the style of Elvis Presley. After that, 3 new samples were generated using the first six seconds of Elvis Presley - Jailhouse Rock as the input signal. At the end, these new songs were analyzed and compared.