eduzhai > Applied Sciences > Engineering >

MusiCoder A Universal Music-Acoustic Encoder Based on Transformers

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 12 pages

Abstract: Music annotation has always been one of the critical topics in the field ofMusic Information Retrieval (MIR). Traditional models use supervised learningfor music annotation tasks. However, as supervised machine learning approachesincrease in complexity, the increasing need for more annotated training datacan often not be matched with available data. In this paper, a newself-supervised music acoustic representation learning approach named MusiCoderis proposed. Inspired by the success of BERT, MusiCoder builds upon thearchitecture of self-attention bidirectional transformers. Two pre-trainingobjectives, including Contiguous Frames Masking (CFM) and Contiguous ChannelsMasking (CCM), are designed to adapt BERT-like masked reconstructionpre-training to continuous acoustic frame domain. The performance of MusiCoderis evaluated in two downstream music annotation tasks. The results show thatMusiCoder outperforms the state-of-the-art models in both music genreclassification and auto-tagging tasks. The effectiveness of MusiCoder indicatesa great potential of a new self-supervised learning approach to understandmusic: first apply masked reconstruction tasks to pre-train a transformer-basedmodel with massive unlabeled music acoustic data, and then finetune the modelon specific downstream tasks with labeled data.

Please select stars to rate!


0 comments Sign in to leave a comment.

    Data loading, please wait...