The article by Julia Barnett, Hugo Flores Garcia, and Bryan Pardo from Northwestern University presents a method to identify similar pieces of music in generative music models. The authors focus on understanding the training data attribution and the influences that shape the creative outputs of these models. They compare the application of CLMR and CLAP embeddings to measure similarity in a set of 5 million audio clips used to train VampNet, a generative music model. The authors suggest their work is foundational to incorporating automated influence attribution into generative modeling, which could help users transition from ignorant appropriation to informed creation.
Publication date: 25 Jan 2024
Project Page: https://tinyurl.com/exploring-musical-roots
Paper: https://arxiv.org/pdf/2401.14542
