Nonparallel Training of Exemplar-Based Voice Conversion System Using INCA-Based Alignment Technique

Hitoshi Suda, Gaku Kotani, Daisuke Saito


This paper proposes a new voice conversion (VC) framework, which can be trained with nonparallel corpora, using non-negative matrix factorization (NMF). While nonparallel VC frameworks have already been studied widely, the conventional frameworks require huge background knowledge or plenty of training utterances. This is because of difficulty in disentanglement of linguistic and speaker information without a large amount of data. This work tackles the problem by utilizing NMF, which can factorize acoustic features into time-variant and time-invariant components in an unsupervised manner. To preserve linguistic consistency between source and target speakers, the proposed method performs soft alignment between the acoustic features of the source speaker and the exemplars of the target speaker. The method adopts the alignment technique of INCA algorithm, which is an iterative method to obtain alignment of nonparallel corpora. The results of subjective experiments showed that the proposed framework outperformed not only the NMF-based parallel VC framework but also the CycleGAN-based nonparallel VC framework. The results also showed that the proposed method achieved high-quality conversion even if the number of training utterances for the source speaker was extremely limited.


 DOI: 10.21437/Interspeech.2020-1232

Cite as: Suda, H., Kotani, G., Saito, D. (2020) Nonparallel Training of Exemplar-Based Voice Conversion System Using INCA-Based Alignment Technique. Proc. Interspeech 2020, 4681-4685, DOI: 10.21437/Interspeech.2020-1232.


@inproceedings{Suda2020,
  author={Hitoshi Suda and Gaku Kotani and Daisuke Saito},
  title={{Nonparallel Training of Exemplar-Based Voice Conversion System Using INCA-Based Alignment Technique}},
  year=2020,
  booktitle={Proc. Interspeech 2020},
  pages={4681--4685},
  doi={10.21437/Interspeech.2020-1232},
  url={http://dx.doi.org/10.21437/Interspeech.2020-1232}
}