Large-Vocabulary Chord Recognition Based on Contrastive Learning and Noisy Student (2023)

Automatic chord recognition can improve the performance of music information retrieval, thereby optimizing the user experience of music electronic products. In order to improve the accuracy of automatic chord recognition, we carry out our research from two perspectives: model structure and training strategies. Firstly, we propose a convolutional neural network with self-attention mechanism as the backbone network. This network well combines the short-time feature capture ability of convolutional neural network with the long-term feature capture ability of self-attention mechanism, to exact better chord features Furthermore, a semi-supervised automatic chord recognition algorithm is introduced based on contrastive learning and a noisy-student network to address the challenge of insufficient high-quality labeled data. The algorithm comprises two stages, namely representation learning and semi-supervised learning. In representation learning stage, we designed a more refined method to construct for positive and negative sample pairs so as to better learn the characteristics of the entire data set. In the context of representation learning, the semi supervised learning stage utilizes both labeled data and high-quality unlabeled data to enhance the training of the model to improve its performance. The experimental results show that the proposed algorithm significantly improves the accuracy of chord recognition.