StemgenRT 0.7.0 separates stereo audio into Drums, Bass, Vocals and Other using three independently trained models. Other is the input minus those three estimates. The plug-in retains the StemgenRT name, host IDs and controls. This release replaces the previous joint separation model.
Download the release for Apple Silicon macOS 14+ (AU/VST3) or Windows x86-64 (VST3). The model and ONNX Runtime 1.26.0 are bundled. macOS packages are ad-hoc signed, not notarized. Close the DAW before replacing the complete plug-in bundle, then restart it. Existing projects will use the new model after updating.
Use 44.1 kHz. Output buses are Main, Drums, Bass, Other, Vocals; Main is the complete delayed input. A 128-sample host buffer gives 256 samples / 5.805 ms of graph-plus-worker latency, excluding hardware latency. Larger or irregular buffers include their additional scheduling reserve in reported latency.
The listening comparison favored cleaner vocals, better silence behavior and less vocal bleed in drums and bass. Other can contain stronger residual errors. Physical M4/M4 Pro sustained timing remains unmeasured for these weights. Correctness tests and algorithmic latency do not establish the CPU deadline. The audio callback never waits for inference; missed output uses aligned fallback. See model details and M4 testing.
On an Apple Silicon Mac, use native arm64 CMake/Ninja and Xcode command-line tools:
git lfs pull
./scripts/download-onnxruntime.sh
cmake --preset release -DCMAKE_OSX_ARCHITECTURES=arm64
cmake --build --preset release --parallel 4
ctest --preset release
./scripts/install-plugins.sh --releaseThe installer replaces complete bundles and signs embedded libraries before the
outer bundle. For Windows, use scripts/download-onnxruntime.ps1, configure with
CMake/Visual Studio, build and test Release, then use scripts/install-plugins.ps1.
The source contains the actual model via Git LFS and independent PyTorch fixtures.
The self-contained FP32 ONNX graph has 14 inputs and 14 outputs: audio plus 13
persistent states. Five belong to Drums, four to Bass and four to Vocals.
The full interface and authenticated EMA identities are under model/.
Matching model, training, objectives, data augmentation, optimizer/EMA recovery,
evaluation and export code is maintained in
StemgenRT-5.8, with a
matching source archive at model/training-source/stemgenrt-0.7.0.tar.gz.
Private experiment archives and draft papers are excluded from this release.