eduzhai > Applied Sciences > Engineering >

Data Cleansing with Contrastive Learning for Vocal Note Event Annotations

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 8 pages

Abstract: Data cleansing is a well studied strategy for cleaning erroneous labels indatasets, which has not yet been widely adopted in Music Information Retrieval.Previously proposed data cleansing models do not consider structured (e.g. timevarying) labels, such as those common to music data. We propose a novel datacleansing model for time-varying, structured labels which exploits the localstructure of the labels, and demonstrate its usefulness for vocal note eventannotations in music. Our model is trained in a contrastive learning manner byautomatically creating local deformations of likely correct labels. Our modelis trained in a contrastive learning manner by automatically contrasting likelycorrect labels pairs against local deformations of them. We demonstrate thatthe accuracy of a transcription model improves greatly when trained using ourproposed strategy compared with the accuracy when trained using the originaldataset. Additionally we use our model to estimate the annotation error ratesin the DALI dataset, and highlight other potential uses for this type of model.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×