Vits-Server 🔥

⚡ A VITS ONNX server designed for fast inference, supporting streaming and additional inference settings to enable model preference settings and optimize performance.

🧪 Experimental purposes only

This project is for experimental purposes only.

If you are looking for a production-ready TTS implementation, go to https://github.com/RVC-Boss/GPT-SoVITS

Advantages 💪

• Long Voice Generation, Support Streaming. 长语音批次推理合并。
• Automatic language type parsing for text, eliminating the need for language recognition segmentation. 自动识别语言类型并处理一切。
• Supports multiple audio formats, including ogg, wav, flac, and silk. 多格式返回写入。
• Multiple models, streaming inference. 多模型初始化。
• Additional inference settings to enable model preference settings and optimize performance. 额外的推理设置，启用模型偏好设置。
• Auto Convert PTH to ONNX. 自动转换pth到onnx。
• Support for multiple languages, including Chinese, English, Japanese, and Korean. 多语言多模型合并支持（任务批次分发到不同模型）。

API Documentation 📖

We offer out-of-the-box call systems.

• Python SDK
• JavaScript SDK

client = VITS("http://127.0.0.1:9557")
res = client.generate_voice(model_id="model_01", text="你好，世界！", speaker_id=0, audio_type="wav",
                            length_scale=1.0, noise_scale=0.5, noise_scale_w=0.5, auto_parse=True)
with open("output.wav", "wb") as f:
    for chunk in res.iter_content(chunk_size=1024):
        if chunk:
            f.write(chunk)

Running 🏃

We recommend using a virtual environment to isolate the runtime environment. Because this project's dependencies may potentially disrupt your dependency library, we recommend using pipenv to manage the dependency package.

Config Server 🐚

Configuration is in .env, including the following fields:

VITS_SERVER_HOST=0.0.0.0
VITS_SERVER_PORT=9557
VITS_SERVER_RELOAD=false
# VITS_SERVER_WORKERS=1
# VITS_SERVER_INIT_CONFIG="https://....json"
# VITS_SERVER_INIT_MODEL="https://.....pth or onnx"

or you can use the following command to set the environment variable:

export VITS_SERVER_HOST="0.0.0.0"
export VITS_SERVER_PORT="9557"
export VITS_SERVER_RELOAD="false"
export VITS_DISABLE_GPU="false"

VITS_SERVER_RELOAD means auto restart server when file changed.

Running from pipenv 🐍 and pm2.json 🚀

apt-get update &&
  apt-get install -y build-essential libsndfile1 vim gcc g++ cmake
apt install python3-pip
pip3 install pipenv
pipenv install  # Create and install dependency packages
pipenv shell    # Activate the virtual environment
python3 main.py # Run
# then ctrl+c exit

apt install npm
npm install pm2 -g
pm2 start pm2.json
# then the server will run in the background

and we have a one-click script to install pipenv and npm:

curl -LO https://raw.githubusercontent.com/LlmKira/VitsServer/main/deploy_script.sh && chmod +x deploy_script.sh && ./deploy_script.sh

Building from Docker 🐋

we have docker pull sudoskys/vits-server:main to docker hub.

you can also build from Dockerfile.

docker build -t <image-name> .

where <image-name> is the name you want to give to the image. Then, use the following command to start the container:

docker run -d -p 9557:9557 -v <local-path>/vits_model:/app/model <image-name>

where <local-path> is the local folder path you want to map to the /app/model directory in the container.

Model Configuration 📁

In the model folder, place the model.pth/ model.onnx and corresponding model.json files. If it is .pth, it will be automatically converted to .onnx!

you can use .env to set VITS_SERVER_INIT_CONFIG and VITS_SERVER_INIT_MODEL to download model files.

VITS_SERVER_INIT_CONFIG="https://....json"
VITS_SERVER_INIT_MODEL="https://.....pth?trace=233 or onnx?trace=233"

model folder structure:

.
├── 1000_epochs.json
├── 1000_epochs.onnx
├── 1000_epochs.pth
├── 233_epochs.json
├── 233_epochs.onnx
└── 233_epochs.pth

Model ID is 1000_epochs and 233_epochs.

when you put model files in the model folder, you should restart the server.

Model Extension Design 🔍

You can add extra fields in the model configuration to obtain information such as the model name corresponding to the model ID through the API.

{
  //...
  "info": {
    "name": "coco",
    "description": "a vits model",
    "author": "someone",
    "cover": "https://xxx.com/xxx.jpg",
    "email": "[email protected]"
  },
  "infer": {
    "noise_scale": 0.667,
    "length_scale": 1.0,
    "noise_scale_w": 0.8
  }
  //....
}

infer is the default(prefer) inference settings for the model.

info is the model information.

How can I retrieve these model information?

You can access {your_base_url}/model/list?show_speaker=True&show_ms_config=True to obtain detailed information about model roles and configurations.

TODO 📝

• Test Silk format
• Docker for automatic deployment
• Shell script for automatic deployment

Acknowledgements 🙏

We would like to acknowledge the contributions of the following projects in the development of this project:

• MoeGoe: https://github.com/CjangCjengh/MoeGoe
• vits_with_chatbot: https://huggingface.co/Mahiruoshi/vits_with_chatbot
• vits: https://huggingface.co/spaces/Plachta/VITS-Umamusume-voice-synthesizer
• espnet: https://github.com/espnet/espnet_onnx
• onnxruntime: https://onnxruntime.ai/