lingxue

lingxue

向着遥不可及的梦想,进发!
steam
youtube
keybase
twitter

Use replay to generate AI singing with one click.

Preface#

Changed the title

Although it is said to be a rescue, in fact, the title is a bit sensationalized

Because I really can't think of any words

In fact, the steps are still UVR processing for accompaniment separation and echo removal, and RVC processing.

But I think most of the problems still occur in the separation, noise, and muffled sound are mostly due to this problem.

The distortion and pitch change are due to the significant difference between the training data and the generated data, which is difficult to solve. Even after training several times, it remains the same.

And using this software directly, it is impossible to fix the problem of noise completely.

Harvest can restore the original sound better, but the effect will be poor if the singer's voice is significantly different.

RMVPE, on the other hand, although the singing style is similar to the input audio, it does not have the problem of pitch deviation in electronic music when the singer's voice is significantly different.

The software does not support AMD graphics cards at all, only CPU.

Preview#

Actually, the middle part has better effects.

It shouldn't have been included, but I accidentally uploaded it and couldn't find a way to delete it, so I just released it.

There is no way to remove the noise part... If the original method is used, it can be removed by importing the AU file...

https://www.tryreplay.io/share/sz3q4

Installation#

Screenshot 2023-12-22 214456

Visit https://www.tryreplay.io/ to download, and then proceed to the next step.

After installation, you have to wait a long time to download the model.

Usage#

After installation, the default settings are like this.

Screenshot 2023-12-22 213245

You can download the model from https://weights.gg/.

Then,

You need to import the model, select
Screenshot 2023-12-22 213429

Import the .pth and .index files.

If it is trained with RVC, the .pth and .index files are not in the same folder. You need to right-click on the model.
Screenshot 2023-12-22 213552

Then copy them in and rename the two files to be the same.

Configuration Section#

The following configuration will yield good results.

If you encounter the problem of pitch deviation in electronic music, you can change F0 to harvest, but it will cause hoarseness and pitch change.

Screenshot 2023-12-22 212835

Conclusion#

It can only be said that it is a very convenient program.

No need to worry about opening two programs and running out of graphics memory.

The problem is, it cannot fix the sound.

Loading...
Ownership of this post data is guaranteed by blockchain and smart contracts to the creator alone.