eduzhai > Applied Sciences > Engineering >

Improving on-device speaker verification using federated learning with privacy

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 5 pages

Abstract: Information on speaker characteristics can be useful as side information inimproving speaker recognition accuracy. However, such information is oftenprivate. This paper investigates how privacy-preserving learning can improve aspeaker verification system, by enabling the use of privacy-sensitive speakerdata to train an auxiliary classification model that predicts vocalcharacteristics of speakers. In particular, this paper explores the utilityachieved by approaches which combine different federated learning anddifferential privacy mechanisms. These approaches make it possible to train acentral model while protecting user privacy, with users data remaining ontheir devices. Furthermore, they make learning on a large population ofspeakers possible, ensuring good coverage of speaker characteristics whentraining a model. The auxiliary model described here uses features extractedfrom phrases which trigger a speaker verification system. From these features,the model predicts speaker characteristic labels considered useful as sideinformation. The knowledge of the auxiliary model is distilled into a speakerverification system using multi-task learning, with the side information labelspredicted by this auxiliary model being the additional task. This approachresults in a 6 relative improvement in equal error rate over a baselinesystem.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×