Updated time: 2021-10-01 18:38:20
05T-1100925
2021-10-01
Know-how
Voice conversion (VC) aims to convert the speech from a source to that of a target without changing the linguistic content. The most typical task is speaker voice conversion. Most traditional methods require parallel training data, that is, the source and target speakers read the same transcript for training. Since such data is difficult to collect, non-parallel training has long been one of the ultimate goals of VC. Through the self-supervised training of the variational autoencoder, non-parallel training can be realized. The integration of generative adversarial networks and cross-domain features can further improve the output speech quality.
- The variational autoencoder can be pre-trained with a large amount of speech
- Output speech quality can be improved by integrating generative adversarial networks
- Output speech quality can be improved by combining cross-domain features
- speaker voice conversion
- impaired speech to normal speech conversion
- voice conversion in role-playing games

Fig1. Variational autoencoder-based voice conversion. x is the input feature, X̂ is the output feature, z is the latent code, and y is the speaker code.
Hsin-Min Wang, Yu Tsao
If you are interested in this technology, please fill in the information on the webpage under the manufacturer option (Request for Technology Licensing, Material Transfer, or Sponsored Research) on the website of this office

Variational Autoencoder-based Voice Conversion