Batch ML
Explicit Process / Separate. ONNX in MLWorker. Sliders do not retrigger it.
Privacy-first local processing
Simple on the surface.
Strict underneath.
One forward STFT. In-place spectral work. One inverse STFT. Make the spectral invariant visible as the product’s technical trust story. Audio is decoded locally; batch ML never runs from a slider.
Input
Upload-only file pick. No microphone.
Decode / resample
Local decode to 48 kHz buffers.
Stem separation
Batch ONNX in a worker (BS-RNN default).
Spectral processing
Fused masks on one STFT representation.
Reconstruction
One inverse STFT. No extra phase passes.
Live-Mix
Zero-ML playback graph: gains, EQ, gate, de-esser.
Export
Local encode. Optional user-initiated Drive save.
Why one spectral pass matters
Fewer transforms = fewer artifacts.
Compatible spectral operations share the same representation instead of repeatedly rebuilding the signal. Live-Mix never starts inference.
Old patternSTFT × N
VoiceIsolateSTFT × 1
Live-Mix
AudioParam + Gate/De-Esser worklets on already-separated stems. Immediate, reversible.
Privacy
No cloud inference. Optional Drive I/O is user-initiated file transfer only.