LARS2.0 is an advanced application for music source separation and drums demixing. It builds upon the foundation of LARS1.0, enhancing its capabilities to handle complete musical tracks.
The implementation of LARS2.0 involves integrating pre-trained models such as HTDemucs4 FT Drums and LarsNet (from LARS1.0) into the C++ codebase.
LARS2.0 can operate in two modes: the same drum-focused mode (separate a stereo drum track into five audio stems: kicks, snare, toms, hi-hat, and cymbals.) as LARS1.0 and a full-featured plugin mode that can separate complete musical tracks, because of that we added 6th stem drums.
It is based on LARS and Music-Source-Separation-Training,
- Clone this repo
- In the
CMakeLists.txt
file, modify the following lines by typing in the path to your LibTorch and JUCE folders:add_subdirectory(/Path/../JUCE ./JUCE)
- On your terminal, go to the project folder and run:
cmake -B build .
- Or build the project with VS code.
Full project with trained models
LarsNet, whose weights are distributed under a CC BY-NC 4.0 license.