eduzhai > Applied Sciences > Engineering >

Deep Self-Supervised Hierarchical Clustering for Speaker Diarization

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 5 pages

Abstract: The state-of-the-art speaker diarization systems use agglomerativehierarchical clustering (AHC) which performs the clustering of previouslylearned neural embeddings. While the clustering approach attempts to identifyspeaker clusters, the AHC algorithm does not involve any further learning. Inthis paper, we propose a novel algorithm for hierarchical clustering whichcombines the speaker clustering along with a representation learning framework.The proposed approach is based on principles of self-supervised learning wherethe self-supervision is derived from the clustering algorithm. Therepresentation learning network is trained with a regularized triplet lossusing the clustering solution at the current step while the clusteringalgorithm uses the deep embeddings from the representation learning step. Bycombining the self-supervision based representation learning along with theclustering algorithm, we show that the proposed algorithm improvessignificantly 29 relative improvement) over the AHC algorithm with cosinesimilarity for a speaker diarization task on CALLHOME dataset. In addition, theproposed approach also improves over the state-of-the-art system with PLDAaffinity matrix with 10 relative improvement in DER.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×