✕

Original fake speech samples - $\large G_{0}$
MELGAN

WAVEGAN

Fake speech samples from user-end models - $\large G_{\phi}$
MELGAN - Non-robust version

WAVEGAN - Non-robust version

Post-processed original fake speech samples Robust Training against adversarial post-processes
Noise
MELGAN - $\large T(G_{0})$ MELGAN - Robust version - $\large G_{\phi}^{R}$

WAVEGAN - $\large T(G_{0})$ WAVEGAN - Robust version - $\large G_{\phi}^{R}$

Gain
MELGAN - $\large T(G_{0})$ MELGAN - Robust version - $\large G_{\phi}^{R}$

WAVEGAN - $\large T(G_{0})$ WAVEGAN - Robust version - $\large G_{\phi}^{R}$

Speed
MELGAN - $\large T(G_{0})$ MELGAN - Robust version - $\large G_{\phi}^{R}$

WAVEGAN - $\large T(G_{0})$ WAVEGAN - Robust version - $\large G_{\phi}^{R}$

Combined Passfilter
MELGAN - $\large T(G_{0})$ MELGAN - Robust version - $\large G_{\phi}^{R}$

WAVEGAN - $\large T(G_{0})$ WAVEGAN - Robust version- $\large G_{\phi}^{R}$