Keywords: Generative Deep Learning, Multi-modal Generation, Creative AI, Self-supervised Representation Learning, Music Generation, Graphic Scores
TL;DR: In this study, we present a tool that can be used to compose new music multi-modally based on user-prompted graphical scores in the style of Gyorgy Ligeti’s Artikulation using self-supervised representation learning.
Abstract: Graphic scores are powerful and expressive symbolic music notations which are promising for music generation in multi-modal settings. However, it is a challenging task to decipher the relationship between the graphic scores and their corresponding musical pieces to explicitly use the creative mapping between them in generative settings. In this work, we connect graphic score and audio domains using self-supervised representation learning to reveal the mapping between these modalities, and utilise this technique to compose music using graphic scores in the universe of György Ligeti's Artikulation, which is a well-known electronic piece. To experiment with and disseminate this approach, we have built an interactive web application in Hugging Face Spaces, designed using the Gradio SDK.
Submission Type: archival
Presentation Type: online
Presenter: Berker Banar