Music Emotion Recognition Using Ensemble Learning

Music has been a part of human experience for centuries and continues to serve not only as entertainment but also as therapy. As music evolves through generations, new genres and subgenres emerge, making it a highly subjective experience. Despite the challenges in determining the emotional content of music, technological advancements offer promising possibilities. This study aims to enhance the accuracy of Music Emotion Recognition (MER), targeting an improvement from the standard $\mathbf{7 3. 1 \%}$ accuracy achieved by previous models. To accomplish this, an ensemble model is proposed, combining three machine learning architectures: (1) ResNet50, (2) DenseNet121, and (3) VGG16. The study utilizes the DEAM dataset, which comprises 1,802 songs (including 58 full-length tracks and 1,744 excerpts of 45 seconds) from various Western popular music genres (e.g., rock, pop, electronic, country, jazz). Each piece is characterized by two numerical values representing Valence and Arousal, based on Robert Thayer’s traditional circumplex mood model, which classifies emotions into four categories: (1) Happy, (2) Calm, (3) Angry, (4) Sad, and four additional emotions. The model’s performance will be evaluated using accuracy, precision, recall, and F1-score metrics. The proposed model obtained an accuracy score of $86.13 \%$, a precision score of $85 \%$, a recall score of $\mathbf{8 4 \%}$ and an $\mathbf{F 1 -}$ score of $\mathbf{8 4 \%}$. The results show that the proposed ensemble model was able to better classify the emotion in music than the individual CNN models. Keywords—circumplex mood model, DEAM dataset, DenseNet121, ensemble model, machine learning, music emotion recognition, ResNet50, VGG16, Western popular music genres.