Improving Music Emotion Recognition by Leveraging Handcrafted and Learned Features

Abstract

Music Emotion Recognition was dominated by classical machine learning, which relies on traditional classifiers and feature engineering (FE). Recently, deep learning approaches have been explored, aiming to remove the need for handcrafted features by automatic feature learning (FL), albeit at the expense of requiring large volumes of data to fully exploit their capabilities. A hybrid approach fusing information from handcrafted and learned features was previously proposed, outperforming separate FE and FL approaches on the 4QAED dataset (900 audio clips). The results suggested that, in smaller datasets, FE and FL could complement each other rather than act as competitors. In the present study, these experiments are extended to the larger MERGE dataset (3554 audio clips) to analyze the impact of the significant increase in data. The best obtained results, 77.62% F1-score, continue to surpass the standalone FE and FL paradigms, reinforcing the potential of hybrid approaches