Aria - A VAE Model using Spectrogram for Music

Abstract

Music travels in the form of sound waves and is incorporated into every aspect of day to day life through entertainment, advertisement, games etc., Music includes pitch, rhythm, dynamics, timbre and texture in its composition. With the rise in technology music can be created with the help of electronic devices to create new and unusual types of music. Even though music is available easily, when it comes to customizing music according to personal taste, a person without any knowledge in music composition cannot make music that is copyright free. We can solve this by deploying a Deep Learning (DL) model where the user can give certain inputs and generate music. We set out to build a Deep Neural Network (DNN) that would ultimately compose music by understanding music theory and creates something completely new. The architecture used in our research is Variational Auto Encoder (VAE) which is trained using spectrogram.