Adversarial DNN-Based Voice Conversion

1. Natural (source): Natural speech of source speaker.
2. Natural (target): Natural speech of target speaker.
3. MGE-FF: Minimum generation error training with Feed-Forward neural networks.
4. MGE-HW: Minimum generation error training with highway networks.
5. ADV-FF: Adversarial training with Feed-Forward neural networks.
6. ADV-HW: Adversarial training with highway networks.

• Yuki Saito, Shinnosuke Takamichi, Hiroshi Saruwatari, "Evaluation of DNN-Based Voice Conversion Deceiving Anti-spoofing Verification," IEICE Technical Report, SP2016-69, vol. 116, no. 414, Jan., 2017. (in Japanese, Student Poster Award) (PDF, Poster)
• Yuki Saito, Shinnosuke Takamichi, Hiroshi Saruwatari, "Adversarial DNN-based voice conversion based on spectral differentials using highway networks," Proc. ASJ, Spring meeting, 1-6-14, 235--236, Mar., 2017. (in Japanese) (PDF, Slide)

Samples	Natural (source)	Natural (target)	MGE-FF	MGE-HW	ADV-FF	ADV-HW
1
2
3
4
5