Table 3 Objective evaluation results of the ablation study on architecture in unseen-to-unseen conversion scenario. “AGAIN-VC” represents the network has neither U2-Net structure nor SaAdaIN. “U2-VC” represents the network has both U2-Net structure and SaAdaIN
From: U2-VC: one-shot voice conversion using two-level nested U-structure
MCD (dB) | Predicted MOS by NISQA | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
SF2TF | SF2TM | SM2TF | SM2TM | Average | SF2TF | SF2TM | SM2TF | SM2TM | Average | |
AGAIN-VC | 5.95 | 6.03 | 5.96 | 6.02 | 5.99 | 3.71 | 3.75 | 3.82 | 3.93 | 3.80 |
w/o SaAdaIN, with U2-Net | 6.11 | 6.16 | 6.19 | 6.20 | 6.17 | 3.85 | 3.91 | 3.81 | 3.97 | 3.89 |
w/o U2-Net, with SaAdaIN | 6.01 | 6.02 | 5.96 | 6.01 | 6.05 | 3.88 | 3.74 | 3.83 | 3.89 | 3.84 |
U2-VC | 6.01 | 6.09 | 6.02 | 6.03 | 6.04 | 4.00 | 3.95 | 3.85 | 3.97 | 3.94 |