A minimum inference engine for DiffSinger MIDI-less mode.

1. Install onnxruntime following the official guidance.
2. Install other dependencies with pip install PyYAML soundfile.
3. Download ONNX version of the NSF-HiFiGAN vocoder from here and unzip it into assets/vocoder directory.
4. Download an ONNX rhythm predictor from here and put it into assets/rhythmizer directory.
5. Put your ONNX acoustic models into assets/acoustic directory.
6. Edit configs/default.yaml or create another config file according to your preference and local environment.
7. Run server with python server.py or python server.py --config <YOUR_CONFIG>.

TBD

1. Train with your own dataset or download pretrained checkpoints from here.
2. Export PyTorch checkpoints to ONNX format. See instructions here.