Note: This bibliographic page is archived and will no longer be updated. For an up-to-date list of publications from the Music Technology Group see the Publications list .

Studying the Application of Deep Learning to the Task of Key Estimation

Title Studying the Application of Deep Learning to the Task of Key Estimation
Publication Type Master Thesis
Year of Publication 2016
Authors Meléndez, B.
Abstract In this thesis, we present, analyse and evaluate two approaches to the music information retrieval (MIR) task of audio key estimation using neural networks and basic deep learning techniques. Although the task has been addressed many times before and with a large variety of methods, few attempts have been made using this increasingly popular machine learning option. Through several experiments, we extract conclusions about the limitations of some of the current approaches, about what is relevant to key estimation and about the ways in which deep learning can successfully be applied to estimate key. We also set a few accuracy baselines in order to evaluate their performance and validate these conclusions. Our approaches surpass the template-based methods used in some MIR libraries such as Essentia, but struggle with a powerful machine learning techniques such as SVM. Finally, to put the acquired knowledge to use, we include a proposal for a hypothetical third approach in the future work section, specifying its most important features.
Final publication https://doi.org/10.5281/zenodo.1162661