Hit Song Prediction: Leveraging Low- and High-Level Audio Features

Assessing the potential success of a given song based on its acoustic characteristics is an important task in the music industry. This task has mostly been approached from an internal perspective, utilizing audio descriptors to predict the success of a given song, where either low- or high-level audio features have been utilized separately. In this work, we aim to jointly exploit low- and high-level audio features and model the prediction as a regression task. Particularly, we make use of a wide and deep neural network architecture that allows for jointly exploiting low- and high-level features. Furthermore, we enrich the set of features with information about the release year of tracks. We evaluate our approach based on the Million Song Dataset and characterize a song as a hit if it is contained in the Billboard Hot 100 at any point in time. Our findings suggest that the proposed approach is able to outperform baseline approaches as well as approaches utilizing low- or high-level features individually. Furthermore, we find that incorporating the release year as well as features describing mood and vocals of a song improve prediction results.