Music-induced emotion flow modeling by ENMI Network

Area of Science:

• Computational Musicology
• Affective Computing
• Machine Learning

Background:

• Music's profound ability to evoke emotions is well-established.
• Music Emotion Recognition (MER) seeks to correlate musical characteristics with affective states.
• Existing MER approaches often simplify the dynamic, continuous nature of emotional responses to music.

Purpose of the Study:

• To conceptualize music-emotion mapping as a multivariate time series regression problem.
• To capture the dynamic flow of emotions within the Arousal-Valence space.
• To introduce and evaluate the Efficient Net-Music Informer (ENMI) Network for MER.

Main Methods:

• The ENMI Network integrates Mel-spectrogram features with time series music data.
• A Music Informer model was trained using combined features for emotional sequence prediction.
• The model's regression accuracy was assessed using Root Mean Square Error (RMSE) on multiple datasets.

Main Results:

• Achieved low RMSE values of 0.0440 for arousal and 0.0352 for valence on the DEAM dataset.
• Demonstrated the model's effectiveness across DEAM, Emomusic, and augmented Emomusic datasets.
• Feature ablation and importance analysis confirmed the contribution of Mel-spectrogram features.

Conclusions:

• The ENMI Network provides an effective approach for modeling the temporal dynamics of music-induced emotions.
• The study highlights the significance of integrating spectral and time series features for accurate MER.
• The proposed model establishes a robust framework for understanding music's emotional impact.