Shota/Boy Voice: The RVC model is more suitable for girls' pseudo-little boys

This is a model file generated by the RVC project. Many of the shota voices are fake little boys of girls, so don't ask why my results aren't great.

The dataset I found is quite large, all sourced from the internet. I incorporated some youthful tones to study whether blending two different tones would improve compatibility between treble and bass.

Obviously, the effect is quite good. I tested two different original voices: Shao Luo and Yu Jie. The latter and the former have one deeper, the other thinner.

The original tone of the reasoning is girlish-like Mi-yō, so after reasoning, the tone leans toward a boy's voice. If the voice is more delicate, it will be in the boy-boy voice-like stage.

During training, I only completed 100 rounds, which was too long and not very useful. After testing, the final model performed well in speaking, but its singing performance was slightly worse. In the end, I packaged the 60th round model and the final two models, which is the pth file.

I also want to mention that, since this is ultimately an AI-generated model, in many model generation processes, I have concluded that when reasoning about songs, it is essential to enunciate words accurately. This is why almost no model has a 100% full-range model, because AI cannot accurately control tone and other controls.

Everyone is welcome to discuss the issue of AI models together; these models are for research purposes only.