Archive
IJMSTA - Vol. 6 - Issue 2 - July 2024
ISSN 2612-2146
Pages: 21
Automatic Mode Recognition and Harmonic Analysis with
Unlabeled Data
Authors: Yui Uehara
Categories: Journal
Abstract - An unsupervised learning method for harmonic analysis is introduced. It is based on a neural hidden semi-Markov model, which is an application of the deep latent variable model technique. The technique uses neural networks to estimate the probability distribution that constitutes the model. The use of a neural network approximation allows for flexible design, such as the generation of an arbitrary number of latent variables; this
is applied herein to produce an arbitrary number of modes. Thus, the proposed model is conceptually not limited to standard major and harmonic minor modes, unlike conventional models. Experiments involving J. S. Bach's four-part chorales resulted in a
convergence of the number of modes to two in many settings and modes with major and minor properties. The proposed model is also valuable as an implementation of unsupervised harmonic analyses, which have not been well studied. The harmonic analysis
results of the proposed model were evaluated using existing labeled data. Although the proposed method has yet to perform as well as existing models that used supervised learning and complex rule design, it has the advantage of not requiring expensive labeled data or rule elaboration.
Keywords: Automatic chord recognition, Harmonic analysis, Hidden semi-Markov model, Neural network
Download paper
