Spaces:
Sleeping
Sleeping
File size: 1,532 Bytes
8012d69 61b2bee 8012d69 c159a8c 8012d69 61b2bee 8012d69 61b2bee | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 | ---
title: Vocal Remover
emoji: 🎤
colorFrom: purple
colorTo: pink
sdk: gradio
sdk_version: 6.19.0
app_file: app.py
pinned: false
license: mit
---
# 🎤 Vocal Remover — Kim_Vocal_2 (MDX-Net ONNX)
Separate **vocals** and **instrumental** from any song using the `Kim_Vocal_2.onnx` model
— the best quality-per-byte ONNX vocal separator available.
## What it does
- Upload any audio file (MP3, WAV, FLAC, M4A, OGG…)
- Get back a clean **Vocals Only** stem and an **Instrumental** stem
- See waveform + mel spectrogram visualizations for all three tracks
- View proxy quality metrics (leakage dB, energy ratio)
## Model
| | |
|---|---|
| **Model** | `Kim_Vocal_2.onnx` (MDX-Net / TFC-TDF U-Net) |
| **Size** | ~67 MB |
| **Vocal SDR** | ~8.9 dB on MVSep benchmark |
| **Runtime** | ONNX Runtime (CPU) |
| **Author** | Kimberley Jensen |
## Performance
Running on free-tier CPU hardware. A typical 3-minute song takes **2–4 minutes** to process.
The model is downloaded automatically on first use (~67 MB) and cached for the rest of the session.
## Settings
| Setting | Default | Notes |
|---|---|---|
| Segment Size | 256 | Lower = less RAM; 512 = better quality |
| Overlap | 0.25 | 0.5 = smoother transitions, 2× slower |
| Enable Denoise | ✅ | Reduces post-separation artifacts |
## Updating the Gradio version
If you fork this Space and get a build error, update `sdk_version` in this file
to match the latest available Gradio version on HuggingFace Spaces:
<https://github.com/gradio-app/gradio/releases> |