[Original] All You Need to Do Is Hands: RVC Beginner Model Training Tutorial (Cloud Tutorial)

frontier

Recently Bao Xiaobai used AI to "replicate" his daughter's voice. For the sound section, he used the currently popular AI software RVC to replicate and restore the voice. Additionally, music content creators on video sites sparked the "AI Stefanie Sun" craze, all using AI software to create and restore the sound.

Try it yourself

here we use

AUTODL

Compared to other cloud platforms, Autodl offers a wide variety of graphics cards and is much cheaper than usual. Or, if you don't want to train, you can use other pre-made high-quality models. Here, we recommend Miaoyin Workshop, which offers thousands of free large models and is a relatively comprehensive model website Miaoyin-RVC Sound Cloning Workshop 。

When choosing a graphics card for computing power, the 3080Ti is recommended by default, as it offers a relatively affordable price and computing power.

select the community mirror RVC-first

Note that V4 is recommended for graphics cards above 40, V3 for cards above 300, and not for other V1 and V2 versions.

After creating it, enter the console and click -Jupyterlab

code output within the console

cd /root/Retrieval-based-Voice-Conversion-WebUI && python infer-web.py --port 6006

we access the link provided by the terminal

Detailed function description

Model reasoning

The first step for a trained model is to refresh it so it can display correctly in the inference tone. Convert male to female +12, female to male -12. Other settings can be set as shown above; leave the rest unchanged by default. After completing the settings, click conversion.

Vocals separated in accompaniment

You can follow the above prompts. Pay attention to whether the audio has reverberation before deciding on different models.

Note that dry audio must have no background noise, noise, or any other sounds. Just follow the above settings. Just to note, the number of rounds doesn't have to be higher the better. Let's test it: with good dry audio quality, 50 rounds is acceptable. The input training audio should be at least one minute long, or it will fail. After training, it is saved in the weights folder.

Model fusion

Simply put, it means fusing model A and B to output model C. One important point to note here is the ratio of A to B fusion. If you want to preserve more of the A model's timbre, you can raise it to a ratio above 0.5.

Other notes

Actually, the steps for training an RVC model aren't that complicated. The real challenge is how to handle dry sounds. Most of the dry sounds we download, even after removing background sounds, are still not perfect, leading to all sorts of strange issues during later inference. The most important thing for a trained large model is handling the dry sound part.