The Emotion Vectorization Algorithm (EVA): Automated Music Generation from Imaging and Emotion Inputs

Abstract

Generative AI tools for the creative arts have become increasingly popular over the past few years. Several well-known models, such as ChatGPT and DALL-E, can even produce writing and artwork comparable to those created by human professionals. Thus, it's no surprise that many technology firms, such as OpenAI and Google, have trained models that can create music as well. These state-of-the-art models usually take in an artist or genre, and they output a song corresponding to the received inputs. However, none of these models are designed to generate music according to an \emph{emotional} input, nor are they able to generate their own styles of music (i.e. they are all trained on well-known works).

Because music is designed to target and evoke specific feelings within the listener, we aim to produce a tool that accounts for this emotional aspect. To this end we create EVA, a new type of generative music model. EVA is the first model takes in a quantitative representation of an emotion as input and returns an instrumentalized musical performance that evokes such an emotion as output. Furthermore, without the reliance on past works of well-known composers for training data, EVA produces a unique style of music that is dissimilar to any particular artists.