Thesis (MSc) - Cyprus International University. Institute of Graduate Studies and Research Information System Engineering

Music genre classification remains a fundamental task in the realm of audio signal processing, playing a pivotal role in various applications such as music recommendation systems, content streaming platforms, and automated music tagging services. This thesis delves into the development and evaluation of machine learning models for music genre classification, leveraging advanced audio feature extraction techniques and state-of-the-art classification algorithms.
The primary objective of this research is to explore the effectiveness of machine learning algorithms in classifying music genres. The study capitalizes on the renowned GTZAN dataset, comprising audio files from ten diverse music genres, to extract crucial audio features using the powerful Librosa library. Various audio features including Chroma STFT, RMS, Spectral Centroid, Spectral Bandwidth, and others are extracted to build comprehensive representations of the audio data.
Two prominent machine learning models, Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM), are employed and rigorously evaluated for their efficacy in music genre classification. The evaluation metrics encompass a wide spectrum, including test accuracy, validation accuracy, ROC AUC, precision, recall, and F1-score, providing a comprehensive assessment of the models' performances across different validation splits.
The findings reveal that both CNN and LSTM models exhibit competitive accuracies and predictive capabilities across different validation splits. The CNN model achieved commendable test accuracies and validation accuracies, with notable strengths in discriminating between music genres, particularly at lower validation splits. Conversely, the LSTM model showcased consistent performance across various splits, demonstrating substantial potential for music genre classification.