Music mashup generation and trend prediction using hybrid attention net

Music mashups creatively blend elements from different songs to create new auditory experiences. Many researchers have explored automating the mashup creation process, often relying on “mashability” metric to determine compatible sections between tracks. We introduce a novel approach to replace this metric. In this study, we first present a framework that generates music mashups by exchanging stems between two distinct tracks. With a thorough analysis of the original music before generation, it uses advanced beat synchronization to maintain musical integrity. In our experiment setup, we selected 10 diverse songs from various styles and tempo ranges to generate 360 unique mashup pieces using 4 distinct strategies. To assess these creations, we conducted rigorous listening tests using a purposely designed format. Three separate groups of music enthusiasts were engaged to assess different mashup strategies, ensuring a comprehensive and robust evaluation. Based on the results of these listening tests, we developed a strategy-encoded hybrid attention network. This model was specifically trained to predict the rankings of mashups that share one common song, enabling it to identify the most compatible pairs given a single input song and a selected strategy from the database. The model demonstrates high predictive accuracy, with a Spearman ranking correlation of 0.75 on the test set, closely aligning with the rankings from the listening tests.