Predicting hit songs: multimodal and data-driven approach
The concept of "Hit Song Science" has been developed to research if popularity can be learned effectively from known song features. Significant effort has been made worldwide to approach this task; however, there is little evidence in literature confirming that the commercial success of a music title can be inferred from technical properties of songs. The main flaw of already developed models for popularity prediction is the lack of consideration for human factors and aesthetic judgements, which are qualitative and highly subjective data. Moreover, "Hit Song Science" as a concept contradicts the natural intuitions of any musically trained composer. Music charts do not just reflect play counts but also manifest the effect of social media impact, internet searches or mentions in the news. Thus, what could improve our understanding of why certain songs are more popular than others is a multimodal approach exploring audio features, melodic, harmonic, and timing structures in songs, as well as marketing and demographics data.
C4DM theme affiliation:
Music Informatics, Audio Engineering, Music Cognition