Audio to Multitrack MIDI (Draft)

Testing only: this resource-intensive tool accepts audio clips up to 20 seconds for every stem selection. It turns a full mix into a multi-track MIDI draft using two neural networks: HT-Demucs FT (ONNX) separates the audio into drums, bass, other instruments, and vocals; then Spotify's Basic Pitch (TensorFlow.js) transcribes each stem into MIDI notes. The stems are assembled into a single .mid file with one track per instrument. This is a draft / starting point for human transcription, not a perfect transcription. Expect missing notes, timing drift, and instrument bleed. Import the result into a DAW and clean it up by hand. Runs entirely in Node (no Python, no GPU).