Musical Key Estimation with Unsupervised Pattern Recognition

Musical key estimation is an important area of music information retrieval (MIR) and music informatics that helps musicians and data analysts to more precisely interact with and understand modern, classical, and marginal population music from a data analysis standpoint. We propose a low-complexity elegant solution to the musical key estimation problem that more efficiently estimates the musical key of a piece of music than previously established methods in the literature. We employ an unsupervised minimization technique to determine both the mode and tonic center of a piece of music in contrast to previous methods that employ supervised learning techniques such as multiple classification models, support vector machines (SVMs), hidden Markov models, multinomial linear regressions, artificial neural networks (ANNs), convolutional neural networks (CNNs), and other high-complexity machine learning algorithms that require training in multiple instances to determine the embedded musical keys. Our approach has a larger scope of possible applications as a result of its simplicity and will be more useful for obscure and archival musical data collection and analytics where large amounts of data are not available for the supervised learning techniques in the literature. Our comparative results confirm the comparability of our method in terms of classification accuracy to previously reportedtechniques with two advantages of lower computational complexity and eliminating the need for large training datasets.