A Comparison Study of Deep Learning Methodologies for Music Emotion Recognition.

Area of Science:

• Computer Science
• Music Information Retrieval
• Artificial Intelligence

Background:

• Classical machine learning for Music Emotion Recognition (MER) faces challenges with manual feature engineering.
• Deep learning (DL) offers automatic feature learning from spectral audio data, but requires substantial labeled datasets.
• MER research often struggles with limited high-quality labeled data.

Purpose of the Study:

• To compare the effectiveness of various classical machine learning and deep learning methods in MER.
• To identify optimal approaches for improving MER performance.
• To address the data scarcity issue in MER through comparative analysis.

Main Methods:

• Conducted a comparative study of diverse classical machine learning algorithms.
• Evaluated multiple deep learning architectures, including Dense Neural Networks and Convolutional Neural Networks.
• Utilized an ensemble approach combining Dense Neural Network and Convolutional Neural Network.

Main Results:

• The ensemble model achieved a state-of-the-art F1 score of 80.20%.
• This represents a significant improvement of approximately 5% over existing baseline results.
• Deep learning methods demonstrated strong performance in automatic feature learning for MER.

Conclusions:

• Combining classical and deep learning paradigms, specifically handcrafted features with learned features, is a promising direction for future MER research.
• Ensemble methods integrating different neural network architectures yield superior results.
• Addressing data limitations remains crucial for advancing MER.