This is a FastAPI server to delegate OpenUtau DiffSinger inference requests. For OpenUtau, you will need my modified OpenUtau.
To start the server: python main.py [-d <MODEL_ROOT_DIR (default to .)>]
OpenUtau will send relative model paths (ex. Singers/Kiritan/acoustic.onnx or Singers/Kiritan/dspitch/pitch.onnx) to the server for inference, and the server will search for the model file under its MODEL_ROOT_DIR. So you can replicate the file structure of the Singers folder and move all models to the remote server, and leave other files in OpenUtau folder. (You can also keep a copy in OpenUtau because when the remote server is unreachable, it will fall back to local inference).
You can also copy vocoder.yaml to the remote server alongside with the finetuned HiFiGAN model. But the server also accepts DDSP JIT model. The trick is to name the JIT file as specified in the vocoder.yaml with .jit suffix, and add a new line model_type: jit to the YAML file. Now the server uses DDSP Vocoder!
GET /exists?model_path={model_path}✔️GET /onnx_info/inputs?model_path={model_path}✔️POST /inference✔️
For one, I want to use DDSP vocoders in OpenUtau, and also try to use torch.compile directly on checkpoints.
This is not a packaged repo, including my modified OpenUtau, you are expected to prepare your own Python environment and build OpenUtau. Later I'll see what can be done.
- Original idea (https://github.com/fishaudio/openutau-remote-host)
- OpenUtau modification reference (https://github.com/fishaudio/OpenUtau)