LARS: An Open-Source VST3 Plug-In for Deep Drums Demixing with Pretrained Models

LARS is an open-source cross-platform VST3 and Audio Unit (AU) plug-in written in C++ and designed for separating a stereo drums mixture into five audio stems. At the core of LARS lies a novel bank of parallel U-Nets pretrained using a collection of over 1224 hours of isolated drum clips, making LARS the first neural-network-based plug-in available for drum source separation. Thanks to a competitive real-time ratio and user-friendly interface, LARS emerges as a versatile tool for music producers and engineers alike. These features makes LARS suitable for tasks such as drum replacement, drum loops decomposition, audio restoration, remixing, and remastering. Furthermore, LARS may find its way into contemporary music production workflows by complementing existing music demixing models.