README.md

Stem Extractor VST3

A real-time, AI-powered audio source separation plugin built with PyTorch, ONNX Runtime, and JUCE. This plugin dynamically isolates Vocals, Drums, Bass, and Other instruments from a full mix directly inside your DAW. Currently installs only on macOS.

Plugin Stem Outputs Demo

Spectrogram Comparison (librosa)

The plugin applies STFT to mask for stems.

Train loss

SI-SDR across Full MUSDB18-HQ Test Set

The baseline U-Net achieves a median SI-SDR of 3.5 dB for Vocals and 2.9 dB for Drums. Output stems corresponding to true stems with no signal (< -60dBFS) are not considered, but there are still extreme negative outliers where the model mistakenly leaks audio into the prediction when there is very subtle true-stem sound.

AI Inference