Automated raga recognition in Indian classical music using machine learning techniques

Abstract

 Indian classical music takes various forms depending on what part of India the dynamics behind it comes from. The current work focused on a binary classification problem involving two popular ragas: Yaman and Bhairavi. This approach allows for a more manageable dataset and facilitates effective training of the models. For machine learning (ML) classifiers, I utilized Logistic Regression, Support Vector Machine with a radial basis function (SVM(rbf)), and XGBoost. These classifiers were chosen for their diverse strengths; Logistic Regression provides a straightforward and interpretable model, while SVM(rbf) is effective for capturing complex decision boundaries. XGBoost, known for its robust performance and efficiency, enhances classification results through its boosting techniques. In addition to traditional ML approaches, we have also explored deep learning (DL) techniques by employing Deep Neural Networks (DNNs) and Long Short-Term Memory (LSTM) networks. DNNs are powerful for learning complex representations of the data, while LSTMs are particularly adept at handling sequential data, making them suitable for audio signals that vary over time. This combination of ML and DL techniques allows for a comprehensive evaluation of the classification capabilities for Yaman and Bhairavi, providing insights into the effectiveness of each method in capturing the nuances of these two distinct ragas. Our method achieves an impressive 88.1% accuracy on the full CompMusic Carnatic dataset and reaches 97% accuracy on a 10 Raga subset, positioning it as the state-of-the-art in Raga recognition. Furthermore, the approach supports sequence ranking, enabling efficient retrieval of melodic patterns from a music database, closely matching the input query sequence.