Optimizing the configuration of deep learning models for music genre classification

Area of Science:

• Music Information Retrieval
• Machine Learning
• Deep Learning

Background:

• Music genre categorization is crucial for music retrieval.
• Manual genre classification is time-consuming.
• Existing machine learning methods show discrepancies from optimal performance.

Purpose of the Study:

• To develop a novel, accurate music genre forecasting approach using deep learning.
• To improve upon existing music genre classification methodologies.

Main Methods:

• Signal preprocessing and feature extraction using Mel Frequency Cepstral Coefficients (MFCC) and Short-Time Fourier Transform (STFT).
• Application of two Convolutional Neural Networks (CNNs) to analyze MFCC and STFT features separately.
• Hyperparameter optimization for each CNN model using the Black Hole Optimization (BHO) algorithm to minimize training error.
• Combining the outputs of the two CNNs with a SoftMax classifier for final genre determination.

Main Results:

• The proposed approach achieved high classification accuracies: 95.2% on the GTZAN dataset and 95.7% on the Extended-Ballroom dataset.
• Demonstrated superior performance compared to previous music genre classification efforts.

Conclusions:

• The novel deep learning methodology effectively categorizes music genres.
• The combination of MFCC, STFT, optimized CNNs, and BHO offers a robust solution for automated music genre classification.