Music genre classification with modified residual learning and dual neural network

Area of Science:

• Music Information Retrieval (MIR)
• Machine Learning
• Digital Signal Processing

Background:

• Music genre classification is challenging due to its subjective nature and reliance on listener interpretation.
• Traditional music genre systems face efficiency and accuracy limitations.
• Deep neural networks offer potential solutions for these challenges in MIR.

Purpose of the Study:

• To improve music genre classification accuracy using deep learning.
• To propose a novel architecture combining modified residual learning and hybrid convolutional neural networks (CNNs).
• To leverage Mel-Spectrograms as input for enhanced feature extraction.

Main Methods:

• Developed a hybrid CNN architecture incorporating modified residual learning.
• Utilized Mel-Spectrograms, representing human-perceived audio signals, as input features.
• Employed identical CNN layers with varied pooling techniques to extract rich hidden information.

Main Results:

• Achieved 87.80% accuracy on the GTZAN dataset.
• Obtained 68.50% accuracy on the FMA dataset.
• Demonstrated comparable performance against existing state-of-the-art models in music genre classification.

Conclusions:

• The proposed deep learning model effectively enhances music genre classification tasks.
• The hybrid CNN architecture with residual learning shows significant potential for MIR applications.
• Mel-Spectrograms provide a robust input representation for accurate music genre identification.