Mygf-f048k usage documentation

Updated on 2025/12/01

This time, we released the Spin version of the base model, which will speed up model convergence to some extent and optimize mute to some extent.

Note: Since the Spin team only uploaded the base mold file during packaging, please prioritize downloading the Contentvec version of the backing film when replacing 48k.json or rmvpe.pt the Contentvec version.

Preface

This pre-trained model completely skips the default base RVC_v2 is trained from scratch. The default RVC_v2 model speaker is 109 bits and is trained on a 50-hour English dataset. The Mygf-f048K is based on the contentvec embedding model and is pretrained using datasets in Chinese, English, and Japanese, with a language dataset of about 100 hours, combined with pretrained RMVPE treble guidance.

Replace

It is worth noting that this model adds speakers, so you need to replace the original RVC_v2 film profile.

configs\v2\48k.json

If you downloaded the official open-source project and the version is V1006, it will definitely be in this location. Before that, you can back up the file and replace it.

Second, the RMVPE model needs to be replaced.

assets\rmvpe\rmvpe.pt

The training model must replace these two files; otherwise, training may fail or the model will become hoarse.

Parameters

The parameters of this backing film are twice as high as RVC_v2. The pre-training parameters are as follows:

• speakers: 308
• languages: Chinese/American/Japanese
• step：≈1.2M
• Training period: 24 days

Detailed tutorial

The directory for the file packaging is as follows

1. rmvpe.pt
2. 48k.json
3. D_Mygf-f048K.pth
4. G_Mygf-f048K.pth

Extract files 1 and 2 to the designated locations as they were replaced in Part 2. You can place files 3 and 4 in any accessible directory. At this point, the new backsheet is deployed.

During training, please first select [Target Sampling Rate - >48K], then modify []. and the path to [loading pre-training base D path].

(modify backing film path).

During training, console output:

DEBUG:infer.lib.infer_pack.models:gin_channels: 256, self.spk_embed_dim: 308

This means the backsheet is correctly identified and loaded successfully. By this point, all steps have been completed.

Solution for RVC error [109, xxxx].

Many users may see similar scenarios on consoles during model configuration or training

[109, xxxx]

This error is generally not due to model corruption, but because different versions of RVC read the configuration file differently, resulting in incorrect matching of the speaker dimension.

The official recommended stable version is:

https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/releases/tag/2.2.231006

If this is your first time setting up an environment, it is recommended to prioritize this version as the training base. When downloading, it is recommended to prioritize the Contentvec version, as the configuration files are mainly included in this version; Next, consider the Spin version.

---

Do you have to use version 1006?

Not necessarily.

Version 1006 is currently the most stable and easiest solution. If you are already using other versions of RVC, you can also manually modify the configuration file to resolve the issue, though the steps will be a bit more complicated.

---

Other versions of RVC workarounds

If you encounter an error [109, xxxx] while running the project, you can follow these steps:

1. Open your RVC project directory, for example:

RVC20240604Nvidia

2. Go to the log folder for the corresponding model, for example:

RVC20240604Nvidia\logs\mygf-xxxx

Mygf-xxxx is the name of your training project; the name may vary depending on the project.

3. In this folder, find :

config.json

4. Open the config.json in Notepad and find this line:

{ "spk_embed_dim": 109 } 

Change
5. 109 to 308, as follows
:

{ "spk_embed_dim": 308 } 

6. Save the file, then restart the training project.

---

Notes

This method is mainly suitable for:

• Configuration file replacement failed
• Use the unofficially recommended version of RVC

If you still get [109, xxxx] error after creating a new training project, you need to re-enter the logs folder of the corresponding project and manually modify the parameters in config.json.

Brief summary:

Whenever 109 appears in the error report, check config.json and set to:

{ "spk_embed_dim": 109 } 

Change to:

{ "spk_embed_dim": 308 } 

Save and retrain to resolve the issue.

Advantages

• No mouth spill will appear.
• models trained that support conversion between different corpora (the "大佐" sound appears in Nichizin (using Mygf-f048k) will not appear.
• can be perfectly trained with about 6 minutes of corpus.
• the tone is closer to the original.

Model trained with a new backsheet

https://apis.klrvc.com/mxgf/3488.html

https://apis.klrvc.com/mxgf/3471.html

https://apis.klrvc.com/mxgf/3451.html

https://apis.klrvc.com/mxgf/3432.html

https://apis.klrvc.com/mxgf/3415.html

That's the complete tutorial for deploying new base molds. Just follow the steps in the tutorial.

Finally, a diagram of the backsheet loss is attached

All of the above tests were conducted in version 2.2.231006:https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI/releases/tag/2.2.231006