Abstract
While the application of neural network has made chord recognition systems acquire considerable accuracy on large vocabularies. However, the lack of labeled data still restricts the further improvement of the performance of automatic chord system. To solve this problem, an automatic chord recognition system based on contrastive learning is proposed in this paper. We train the model with a large amount of unlabeled data and labeled data. The application of contrast loss helps the model generate a more uniform representation of feature. We also introduce the teacher-student network to generate pseudo labels for unlabeled data, which are used to screen negative samples. In this way, the damage of “false” negative samples to the model can be reduced. Experimental results show that the proposed method significantly improves the recognition ability and generalization of automatic chord recognition system.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have