please be patient when applying for datasets :)

[MoonInTheRiver]

I will give you the link within seven days. If not, email me again.

[daxiangpanda]

话说能不能给条歌唱的例子数据呢？一条或者几句就对了？ phone，text，textGrid，wav这几个数据都是啥样的？

[keonlee9420]

Thanks @MoonInTheRiver for the note!

I checked that the current code doesn't support tasks.vocoder. But we may have to train/fine-tune them when we apply new datasets. So could you please include the hifigan/parallelwavegan trainer/fine-tuner/inferer in the next release for adapting new datasets?

[MoonInTheRiver]

话说能不能给条歌唱的例子数据呢？一条或者几句就对了？ phone，text，textGrid，wav这几个数据都是啥样的？

see 'dataset preview'

[MoonInTheRiver]

Thanks @MoonInTheRiver for the note!

I checked that the current code doesn't support tasks.vocoder. But we may have to train/fine-tune them when we apply new datasets. So could you please include the hifigan/parallelwavegan trainer/fine-tuner/inferer in the next release for adapting new datasets?

Actually the training codes for vocoder are not assembled even in diffsinger's local codebase. For TTS, people can use the GT-melspectrograms and GT-wavs processed by our codes to train the official implementation of HifiGAN. But remember to replace "y_mel" in "F.l1_loss(y_mel, y_g_hat_mel).item()" by "mel_spectrogram(y, ...)", since if you use the GT-mel processed by our codes, the mel_spectrogram(y_g_hat, ...) will be not consistent with y_mel in terms of scale.

[daxiangpanda]

thank you very much can you share the loss value of your pretrained model? and what kind of gpus your model use when training the pretrained model？

[MoonInTheRiver]

thank you very much can you share the loss value of your pretrained model? and what kind of gpus your model use when training the pretrained model？

Just use the provided pre-trained model, and run additional 100 steps. Then you will get that.