A Fine-tuning Method for Automatic Chord Estimation Using the Foundation Model

In this paper, we examine a method for automatic chord estimation (ACE) using foundational models. We present a series of configurations for a system that estimates chord progressions by fine-tuning a foundation model with self-supervised learning, which addresses the problem of insufficient labeled data in music informatics, and examine how hyperparameters and minor configuration changes in the fine-tuning process affect estimation accuracy. The purpose of this study is to show how an ACE system using a foundation model pre-trained from unlabeled data should be specifically designed and implemented, and to indicate the factors during training that affect its accuracy. In the study, the conditions at learning time were compared using MusicFM, which was published in a previous study that used the foundation model to acquire musical information. The comparison revealed that while the application of techniques in deep learning, such as the use of the Nadam optimizer and the implementation of stepwise fine tuning, improved accuracy, factors that increase the complexity of learning per epoch, such as the weighting of the loss function and the complexity of the output layer, could hinder the improvement in accuracy. Overall, the model performed as well or better than existing models in estimating chord progressions for triads. These findings contribute to the promotion of the use of pre-trained foundational models in ACE, helping to understand the knowledge from the first steps of implementation to further extensions.