Chord Classification Using Machine Learning

Abstract

Chord classification is a fundamental task that enables automated music transcription and analysis in Music Information Retrieval. This paper presents a comparative study of the machine learning approaches for chord recognition from audio signals. The proposed framework deals with two feature extraction methods– Pitch Class Profile (PCP) and spectrograms, which are evaluated across three classification models: Support Vector Machines (SVM), Multilayer Perceptron (MLP), and Convolutional Neural Networks (CNN), and these models are trained to classify 24 chord classes comprising 12 major and minor triads. Upon experiment, the result demonstrates that PCP- based feature extraction significantly out performs spectrogram-based methods, particularly with limited training data. The MLP classifier achieves the highest performance with 0.99 precision, recall, F1-score, and accuracy, though these metrics indicate potential overfitting. The SVM model yields 0.96 accuracy with better generalization capability on unseen data. The CNN model achieves up to 0.97 training accuracy but exhibits poor generalization. The finding reveals that with PCP-enabled method can be used effectively and develop a robust chord classification system. This work highlights the potential of machine learning in improving chord classification accuracy and sets the stage for future development by refining models and expanding the dataset.