README.md

---
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>