Open source project: https://github.com/jerryuhoo/VTuberTalk
Solved countless errors, modified multiple scripts, worked hard for days and nights, and finally completed the project, but the success was unsuccessful. Please note that if the training sample contains less than 500 sentences, the effect will be very poor and you can give up directly.
1. Install
the cpu version: If you are just using it, it is recommended to install this version.
conda create -n paddlespeech python=3.8
conda activate paddlespeech
pip install -r requirements_cpu.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
2. Install the cpu version of paddlepaddle
# The cpu version example requires paddle 2.3.0 or above
python3 -m pip install paddlepaddle==2.4.2 -i https://pypi.tuna.tsinghua.edu.cn/simple
python3 - c "import paddle;paddle.utils.run_check()"
3. Obtain audio.
How to obtain audio from station B: You can use bilibili assistant to download Vtuber's recorded flv file and then convert it to a wav file.
How to get audio from YouTube: You can use the YouTube downloader on TamperMonkey to download the mp4 file and then convert it to a wav file.
Install dependent libraries:
pip install pydub
python tools/video_to_wav.py --path <data to folder or file>
I use the format factory here, which can convert and edit any format, and finally directly output data/