voice-changer icon indicating copy to clipboard operation
voice-changer copied to clipboard

Onnx and Pth Issue

Open benc1221 opened this issue 1 year ago • 18 comments

Issue Type

Bug Report

vc client version number

MMVCServerSIO_win_onnxgpu-cuda_v.1.5.3.9

OS

Windows 11

GPU

RTX 4080

Clear setting

yes

Sample model

no

Input chunk num

yes

Wait for a while

The GUI successfully launched.

read tutorial

yes

Voice Changer type

RVC

Model type

ONNX

Situation

First, eveerytime i use any onnx file with any of the crepe settings, its sounds robotic and sounds like its whispering. In the previous version where i use my index files with .pth, it never made my models sound watery and muffled, but pth works with all crepe options, so the major issue with this version is onnx is not working with any of the crepe options, the index files when you set it above 0.5 it makes the voice sound watery and muffled in both pth and onnx.

benc1221 avatar Jul 08 '23 01:07 benc1221

I have this too, index just sounds terrible on this new version, and all the new crepe options sounds really bad. I've went back to 1.5.3.8a for now.

NataIynn avatar Jul 08 '23 01:07 NataIynn

Are all of the crepe incorrect? "crepe", "crepe_tiny", "crepe_full".

I didn't modify "crepe" from the previous version, so there might be a problem with the other logic.

w-okada avatar Jul 08 '23 08:07 w-okada

none of the crepe options work with the onnx file

benc1221 avatar Jul 08 '23 10:07 benc1221

tell me your onnx file's properties. sample rate, f0, v1 or v2

w-okada avatar Jul 08 '23 10:07 w-okada

48k crepe v2

benc1221 avatar Jul 08 '23 10:07 benc1221

Does the same issue occur with the sample model?

w-okada avatar Jul 08 '23 10:07 w-okada

yes

benc1221 avatar Jul 08 '23 10:07 benc1221

Really??? According to reports from other people, they said that there were no issues with the sample model, but could this depend on the environment as well??

w-okada avatar Jul 08 '23 10:07 w-okada

the sample models only work for harvest and dio mode for me

benc1221 avatar Jul 08 '23 10:07 benc1221

if you can share the output of sample model with this song. This is a famous Japanese children's song.

https://drive.google.com/file/d/1iCErRzCt5-6ftALcic9w5zXWrzVXryIA/view (sorry link is not valid, modified)

w-okada avatar Jul 08 '23 10:07 w-okada

which f0 detector i need to use?

benc1221 avatar Jul 08 '23 10:07 benc1221

crepe

w-okada avatar Jul 08 '23 10:07 w-okada

https://drive.google.com/file/d/1WmGTOi74vVdJk4SSMNHoFtycqcZUE9lW/view?usp=sharing Screenshot 2023-07-08 054825

benc1221 avatar Jul 08 '23 10:07 benc1221

Thanks!! That's really strange. Let's investigate.

w-okada avatar Jul 08 '23 10:07 w-okada

This may be because crepe cannot be inferred unless the voice is longer than a certain length. Is your voice hoarse in the Tsukuyomi-chan model?

nadare881 avatar Jul 08 '23 14:07 nadare881

only when i use crepe, crepe tiny, and crepe_full

benc1221 avatar Jul 08 '23 15:07 benc1221

I released bugfixed version. try v.1.5.3.9a

w-okada avatar Jul 09 '23 03:07 w-okada

https://drive.google.com/file/d/1nkCpNua8SiZr7a5vivmLsTAT0Hzb1637/view?usp=drive_link https://drive.google.com/file/d/1yycJjFV5Bz7eo0DHRY5is1cQaukH8xz9/view?usp=drive_link the issues im now facing with the index file but all crepe options for both onnx and pth are working good

benc1221 avatar Jul 09 '23 15:07 benc1221

OKey,

And at which version you can use index correctly ?

I search diff with that version.

w-okada avatar Jul 10 '23 21:07 w-okada

1.5.3.7

benc1221 avatar Jul 10 '23 22:07 benc1221

no clue...

w-okada avatar Jul 12 '23 23:07 w-okada

no clue...

w-okada avatar Jul 18 '23 00:07 w-okada

no clue and new version released. try it and issue remains, open new issue. sorry.

w-okada avatar Jul 21 '23 14:07 w-okada