Converting Source Separation Models to ONNX for Real Time Usage in DJ Software
Demucs v4 is a state-of-the-art open-source music source separation model developed by Antoine Défossez. While it provides exceptional quality in separating audio into stems, it is currently implemented in PyTorch and therefore cannot be used directly in C++ applications or run efficiently on hardware accelerators through ONNX Runtime. Exporting the model to ONNX allows developers to integrate stem separation directly into their platforms, opening up use cases like live stem manipulation, AI-assisted mixing, and educational visualization tools. Through this talk where I demonstrate exporting and embedding the model in an open source DJ platform, I want to provide a reference for further integration into audio applications.
Anmol Mishra
Research Assistant
Music Technology Group
Anmol is a Research Assistant at the Music Technology Group (MTG) and a masters student at UPF Barcelona, working on learning expressive groove from Afro-Latin drumming for algorithmic rhythm generation. Previously he worked as a Machine Learning Engineer at Samsung’s Computer Vision Group in Seoul. Prior to that, he earned his Bachelors in Computer Science from IIT Bombay. He's been performing as a DJ for the past two years at clubs and festivals in Seoul.