maum-ai/phaseaug

ICASSP 2023 Accepted

phaseaug's Issues

metallic sound when applying phaseaug on avocodo

Hi,

Thanks for your great work, I try to apply phaseaug on avocodo(a littel different from hifigan on discriminator).
But there is metallic sound, now the training steps is around 200k, did you encounter the same problem on hifigan?

Thanks

training error

when I try to use this repo, i got the error like this "RuntimeError: stft input and window must be on the same device but got self on cuda:0 and window on cpu", did i Miss something?

Univnet and PhaseAug

Hello,
First of all thanks for interesting work.

I'm just curious if you tried to combine PhaseAug with mindslab implementation of Univnet (https://github.com/mindslab-ai/univnet)?

Pytorch 2 cannot compile forward_sync

Hi guys,

Thanks for your work on this project.

I used phaseaug in my vocoder which worked well.

but when I compile the whole model in pytorch 2 and get something wrong.

I minify the codes:

phaseaug = PhaseAug()
compiled_phaseaug = torch.compile(phaseaug.forward_sync)
wav = torch.randn([32, 1, 9600])
compiled_phaseaug(wav, wav)

Do you have any ideas about that?

phaseaug on bigvgan

Hi,

Thank you for the work.
I am wondering if phaseaug has been tried on bigvgan which uses MRD (univnet discriminator)? would it be straightfoward to adopt the hifigan example?

Thanks

Missing variable in forward_sync

I got an exception from forward_sync:

...
  File "/usr/local/lib/python3.8/dist-packages/phaseaug/phaseaug.py", line 117, in forward_sync
    phi = self.sample_phi(self, X.shape[0] // 2) #[2B, nfft//2+1]
NameError: name 'X' is not defined

X is the result of stft and is removed in version 1.0.0

about periodicity artifacts

what is periodicity artifacts? can you show me the mel-image?

why add this line

https://github.com/mindslab-ai/phaseaug/blob/07d1c74244fc55359d022bfb1bd21acce2cc9066/train.py#LL175C8-L175C20

I think aug is frozened, not trained.

Will it add new inference cost to the vocoder?

Thanks for the job.
I am trying to find some training-auxiliary component for hifigan, which can help leaning the periority but without drawing any new parameters into the Generator. Is the phaseaug just augmenting trainging-data on phrase? Will it introduce new trainable parameters to the vocoder model?
We are tring to speedup the HiFiGAN model, the decreased structures and parameters will lead to more significant High-Frequency distortion on spectrum. Will the PhaseAug increase the ability to reconstruct the signal on High frequency band？ Want the speech spectrum displayed about comparation before and after PhaseAug. thanks.

where to put this code in VITS?

Hi, where should i put this code in VITS?
Which is the y_g_hat?

Audio samples inaccessible

Hello,

I noticed that the audio samples link given in README.md is not accessible. I guess you have migrated the repository from "github.com/mindslab-ai" to "github.com/maum-ai". Changing the link manually worked.

Changed in version 2.0: Real datatype inputs are no longer supported.

https://pytorch.org/docs/stable/generated/torch.istft.html#:~:text=datatype%20inputs%20are-,no%20longer%20supported,-.%20Input%20must%20now

in PhaseAug -> forward9(_sync)

x_aug = torch.istft(

RuntimeError: istft requires a complex-valued input tensor matching the output from stft with return_complex=True.

Maybe:

X_aug = X * torch.complex(phi.cos(), -phi.sin())