yl4579 / hiftnet Goto Github PK

Hi, thx for your great work!
I wonder is it possible to add multi band to further improve inference speed like https://github.com/MasayaKawamura/MB-iSTFT-VITS

why in train.py，botch loss_disc and loss_gen use "generator_TPRLS_loss"?
loss_disc_s += generator_TPRLS_loss(y_ds_hat_r, y_ds_hat_g)
loss_gen_s += generator_TPRLS_loss(y_ds_hat_r, y_ds_hat_g)

Thanks for your greak work

I got warnings when torch==1.8 and 1.10
（Under 1.8 will return error about return_complex of torch）

UserWarning: The default behavior for interpolate/upsample with float scale_factor changed in 1.6.0 to align with other frameworks/libraries, and now uses scale_factor directly, instead of relying on the computed output size. If you wish to restore the old behavior, please set recompute_scale_factor=True. See the documentation of nn.Upsample for details. 

  "The default behavior for interpolate/upsample with float scale_factor changed "

UserWarning: Default upsampling behavior when mode=linear is changed to align_corners=False since 0.4.0. Please specify align_corners=True if the old behavior is desired. See the documentation of nn.Upsample for details.

  "See the documentation of nn.Upsample for details.".format(mode)

But there is no special version of torch in requirements.txt

Are these warnings as expected?
or
What's the correct version of torch?

Thanks

"for dr, dg" , not "for dg, dr"?

The quality of demo samples seems to be quite good, even in noisy environments or for unseen languages. Did you use the same f0 network (pretrained on LibriTTS) for these out-of-domain experiments? I'm also wondering whether it's okay to use the F0 network as is for unseen languages. If you have any experimental insights, I'd appreciate learning about them.

Hi! I wanted to know if you know about Vocos and if you compared to it since it uses similar principles and has similar results.

Hello, yl4579. In 2.1.1. Efficient Source Generation, you do the We note that the L factor in Equation 6 scales the value by the hop size, as Equation 5 now integrates with 1/L of steps compared to steps in Equation 2.. I want to know what is the performance benefit of this change?