Bass stem generation and swapping for music re-creation

Music mashups combine elements from different songs to create unique auditory experiences. This study builds on previous work with drum stem swapping by incorporating bass stem swapping into mashup generation. Two strategies are evaluated: bass stem-swapped mashups and drum-and-bass stem-swapped mashups. The preference correlations between these strategies are low, averaging 0.33. This result highlights the distinct listening experiences created by swapping different stems within the same song pair. Additionally, harmonic similarities across song pairs show no significant impact on mashup quality. Another contribution of this study is a controllable bass stem generation model guided by other stems. This model efficiently generates bass stems using limited data. It also allows users to control the dynamics and rhythms of the generated bass stems, offering greater creative flexibility. This design fosters human-machine collaboration in the music creation process. Future work will explore additional stem-swapping strategies and develop predictive models for mashup preferences. These advancements aim to enhance tools for automated music re-creation.