- This is a project aiming to create my own neural network inference library.
- This project has two parts, one is the parser part which has not yet been made public and this is the inference part which has been made public but I considered not done, I will try to spend time to make it much better.
- This program here utilized the Dual-signal Transformation LSTM Network(DTLN) to denoise audio comming in from a microphone(with miniaudio full duplex)
- This part is made public first because it can produce its own goal, and I must push my Medium story immediately so I rushed this out.
- The parser is not yet made public, it's purpose is to parse an ONNX model file and create a header file similar to "model1.h", "model1.c", "model2.h", "model2.c".
- The generated header and source file contain the node description and weight of the file, it is then added into the inference part.
- It's dependency is protobuf.
- This project.
- It runs the DTLN model by using the pair of files created by the parser.
- There is still a lot of optimization and making it conform to coding guidelines should be done.