SMU Data Science Review


Music is incorporated into our daily lives whether intentional or unintentional. It evokes responses and behavior so much so there is an entire study dedicated to the psychology of music. Music creates the mood for dancing, exercising, creative thought or even relaxation. It is a powerful tool that can be used in various venues and through advertisements to influence and guide human reactions. Music is also often "borrowed" in the industry today. The practices of sampling and remixing music in the digital age have made cover song identification an active area of research. While most of this research is focused on search and recommendation systems, plagiarism is a real industry wide problem for artists today. Our research seeks to describe a framework of feature analysis to improve cross-similarity, song-to-song, similarity distance measurements. We do this with the context that cover songs represent a fertile training ground to prove methods that can later be applied to plagiarism use cases. Our proposed method preprocesses songs by first source separating the songs into its constituent tracks prior to feature generation. This is otherwise known as "stemming". These subsequent spectral and distance features are then analyzed to provide evidence of improvement in overall modeling and detection performance. We find that using stem distances and overall distance measures achieves an uplift of 61.8% increase in Accuracy, a 59.2% increase in AUC, a 304.7% increase in Precision, and a 105.1% increase in F1 score for a regularized logistic regression. This process can be directly applied to more sophisticated deep learning frameworks.

Creative Commons License

Creative Commons Attribution-Noncommercial 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License