Automatic Chord Estimation Using Generative Adversarial Networks ～Improved DCGAN-based Method, Applied CGAN-based Method～

Abstract

Automatic chord estimation is an important research topic in the field of music analysis. In existing research, a method using DCGAN has been proposed, but the accuracy rate of automatic chord estimation is only 73.2%. Therefore, in this study, we propose an improved method using DCGAN and a method using CGAN. The DCGAN-based method was improved by using tanh functions for all the coupling and output layers to unify the regions, and by adding L2 regularization to the Minibatch Discrimination layer and all the convolution layers to suppress Mode Collapse and overlearning. Experimental results showed a 91.7% correct rate of automatic code estimation. For the CGAN-based method, a one-hot vector was used as the input for the condition, and two methods were tested: one method added to the end of the MIDI data, and the other method replaced the beginning of the MIDI data. The experimental results showed that the correctness rates of the two methods for automatic chord estimation were 90.5% and 90.9%, respectively. This study has made it possible to improve the accuracy rate of the DCGAN-based method in automatic code estimation. Furthermore, it was shown that the method using CGAN also achieved a high accuracy rate.