Analysis of the Originality of Gen-AI Song Audio
Rajesh Fotedar (University of Miami), Tom Collins (University of Miami)*
This paper will be presented in person
We are investigating whether an AI-based generative music (Gen-AI) system returns original song audio in response to the same text prompts. To determine a similarity threshold for our analysis, we compare audio from human-composed song pairs that were previously involved in actual copyright disputes. Participants are tasked with sending the same 8 text prompts, specifically designed for our investigation, to the Gen-AI system. By comparing all returned Gen-AI songs to each other, a maximum 8-beat correlation value is determined for each unique song pairing. Gen-AI songs pairs are deemed too similar -- or unoriginal relative to each other -- if their maximum 8-beat correlation value exceeds the similarity threshold. We observe that 54% of Gen-AI song pairs fall into this category. Prompts that exhibit the highest probability for exceeding the threshold include mentions of a specific musical genre or song title. Our preliminary results suggest that the Gen-AI system struggles to generate original material when different users send the same prompts.