eduzhai > Applied Sciences > Engineering >

Recognition-Synthesis Based Non-Parallel Voice Conversion with Adversarial Learning

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 5 pages

Abstract: This paper presents an adversarial learning method for recognition-synthesisbased non-parallel voice conversion. A recognizer is used to transform acousticfeatures into linguistic representations while a synthesizer recovers outputfeatures from the recognizer outputs together with the speaker identity. Byseparating the speaker characteristics from the linguistic representations,voice conversion can be achieved by replacing the speaker identity with thetarget one. In our proposed method, a speaker adversarial loss is adopted inorder to obtain speaker-independent linguistic representations using therecognizer. Furthermore, discriminators are introduced and a generativeadversarial network (GAN) loss is used to prevent the predicted features frombeing over-smoothed. For training model parameters, a strategy of pre-trainingon a multi-speaker dataset and then fine-tuning on the source-target speakerpair is designed. Our method achieved higher similarity than the baseline modelthat obtained the best performance in Voice Conversion Challenge 2018.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×