eduzhai > Applied Sciences > Engineering >

Semi-supervised learning using teacher-student models for vocal melody extraction

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 8 pages

Abstract: The lack of labeled data is a major obstacle in many music informationretrieval tasks such as melody extraction, where labeling is extremelylaborious or costly. Semi-supervised learning (SSL) provides a solution toalleviate the issue by leveraging a large amount of unlabeled data. In thispaper, we propose an SSL method using teacher-student models for vocal melodyextraction. The teacher model is pre-trained with labeled data and guides thestudent model to make identical predictions given unlabeled input in aself-training setting. We examine three setups of teacher-student models withdifferent data augmentation schemes and loss functions. Also, considering thescarcity of labeled data in the test phase, we artificially generatelarge-scale testing data with pitch labels from unlabeled data using ananalysis-synthesis method. The results show that the SSL method significantlyincreases the performance against supervised learning only and the improvementdepends on the teacher-student models, the size of unlabeled data, the numberof self-training iterations, and other training details. We also find that itis essential to ensure that the unlabeled audio has vocal parts. Finally, weshow that the proposed SSL method enables a baseline convolutional recurrentneural network model to achieve performance comparable to state-of-the-arts.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×