Ai Tanglong/Nai, Tanglong/Milk GPT-SoVITS model

When training this model dataset, loudness is uniformly matched, resulting in a lower sound output. In post-production, simply adjust the volume.

Added AI-generated Japanese data to the original dataset, which works well for Japanese because the Japanese leans toward the Crayon Shin-chan tone, which incorporates the Crayon Shin-chan tone.

The compressed files have a total of:

Processed dataset slicing.

Reference audio.

V2 Two model files.

Last time,

a friend messaged me saying they couldn't connect to drive 123, so I also set up a Google Drive to help those who can't access drive 123 download it. To demonstrate the inference, I used two models in Chinese and one in Japanese to demonstrate this model. After processing the dataset, there is no noise or distortion, and after downloading, it can be trained automatically. I used the default training parameters. But the effect is already very good.

Note: I trained on GPT-SoVITS V2, so you should use GPT-SoVITS V2 for inference; otherwise, using V1 may cause incompatibility issues.

The RVC version is currently in training and may be released a bit later.

This model is for research and study only. Please proactively delete it 24 hours a day. Do not use it for commercial purposes. Any violation will result in your own legal responsibility.