STFT question in multi-frequency discriminator

[jerryuhoo]

VISinger2/modules/stft.py

Line 377 in 6d63887

self.mel_scale = MelScale(n_mels=(fft_size // 2 + 1),

Hi, I found a potential bug. The sample rate is not passed into MelScale in this line in the multi-frequency discriminator, so the default sample rate is 24000. However, the sample rate in the config file is 44100. Is this a bug or it doesn't matter? Also, may I ask did you implement MultiFrequencyDiscriminator yourself, or do you have a reference code? How do you choose the parameters (such as domain="double" and hop lengths)? Thanks!

[zhangyongmao]

Thanks for the correction, the MelScale setting is wrong. The MultiFrequencyDiscriminator implementation refers to another code from our lab. The domain setting to double increases the log-spectrum, but I did not compare the impact of whether to set it to double or not. I usually set the hop size of the MultiFrequencyDiscriminator between 2.5ms and 15ms. Too many discriminators may increase the training time, so I think three to six discriminators are more appropriate. A uniformly distributed hop size is probably the best choice.