eduzhai > Applied Sciences > Engineering >

Guided multi-branch learning systems for sound event detection with sound separation

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 5 pages

Abstract: In this paper, we describe in detail our systems for DCASE 2020 Task 4. Thesystems are based on the 1st-place system of DCASE 2019 Task 4, which adoptsweakly-supervised framework with an attention-based embedding-level poolingmodule and a semi-supervised learning approach named guided learning. Thisyear, we incorporate multi-branch learning (MBL) into the original system tofurther improve its performance. MBL uses different branches with differentpooling strategies (including instance-level and embedding-level strategies)and different pooling modules (including attention pooling, global max poolingor global average pooling modules), which share the same feature encoder of themodel. Therefore, multiple branches pursuing different purposes and focusing ondifferent characteristics of the data can help the feature encoder model thefeature space better and avoid over-fitting. To better exploit thestrongly-labeled synthetic data, inspired by multi-task learning, we alsoemploy a sound event detection branch. To combine sound separation (SS) withsound event detection (SED), we fuse the results of SED systems with SS-SEDsystems which are trained using separated sound output by an SS system. Theexperimental results prove that MBL can improve the model performance and usingSS has great potential to improve the performance of SED ensemble system.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×