Automatic Chord Estimation Using Deep Learning Focused on Overtone Structure

A chord in music is two or more notes played simultaneously, and if the order of these chords can be estimated, the mood of the piece can be determined. However, to manually annotate chords, it is time-consuming to account for musical elements such as the key of the piece. Conventional methods automatically estimate chords using deep learning and prior probability, but the estimation accuracy is low. In this study, we believe that the cause of this misidentification is the tension in the chord. Therefore, we propose a method of attenuating tones that are unnecessary for estimation, based on the relationship between the base tone and the fifth-degree overtone in the chromatic vector of the input feature. To confirm the estimation accuracy, the Variational Autoencoder (VAE) model is trained using a supervised method using Markov prior distribution on annotated data of 1,217 Jpop and Western music songs, post witch the experiments were conducted. Consequently, there was a tendency for the conventional method to have better estimation accuracy than the proposed method.