A Voice Cloning Method Based on the Improved HiFi-GAN Model

Table 3

The key training parameters of the feature prediction network.

Dimensions of the speaker embedding vector	256
Silence duration (s)	0.4
Utterance duration (s)	16
Mel spectrum channel number	80
Initial learning rate	0.003
Final learning rate	0.00005
Spectral window length (ms)	50
Spectral window shift (ms)	12.5