Audio Separator

Upload → pick a use case → get labelled stems, ZIP download. Powered by audio-separator + Gradio.

Input & Use Case
Use Case (one-click preset)

Pick a goal — we'll choose the best model + params.

ℹ️ Removes the lead vocal so you can sing along. Best for pop/rock.
Model (auto-selected, override anytime)
Output format
0.1 1
0.1 1
1 16
Output Stems

Ready. Upload an audio file and click Separate Audio.
Roformer / MDX23C
32 4000
2 50
-12 12

Quick start

  1. Upload or record audio in the input box on the left.
  2. Pick a use case from the dropdown — the right model is auto-selected.
  3. Click 🚀 Separate Audio. The first run downloads the model (can be hundreds of MB).
  4. Listen to each stem, A/B against the input, and download a ZIP of everything.

When to override the model

  • Karaoke → MelBand Roformer | Karaoke V2 by Gabox is a great default.
  • Clean a cappella → BS-Roformer-Viperx-1297 (highest reported SDR).
  • De-reverb → MelBand Roformer | De-Reverb by anvuew.
  • De-noise → Mel-Roformer-Denoise-Aufr33.
  • Multi-stem remix → htdemucs_ft.yaml (4 stems) or htdemucs_6s.yaml (6 stems).
  • CPU-only / speed → UVR-MDX-NET-Inst_HQ_3.onnx.

Advanced parameters cheat sheet

Param Effect
segment_size Larger = better quality, more RAM.
overlap Higher = better quality, slower.
pitch_shift Shifts pitch ±12 semitones during processing (Roformer/MDX23C).
enable_denoise MDX-NET only — removes residual noise on stems.
enable_tta VR Arch — Test-Time Augmentation, slower but cleaner.
shifts Demucs — number of random-shift passes (higher = better, slower).

Tips

  • For best results use lossless inputs (WAV/FLAC).
  • Models auto-cache in Model cache directory — first run is slow, later runs are fast.
  • Use Advanced mode to expose every parameter the upstream HF Space exposes.

Credits
DEVICE CUDA · ENGINE installed · THEME BLACK / SYNE