---
title: Audio Restoration V3
emoji: 👻
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: "4.44.0"
app_file: app.py
pinned: false
license: mit
hardware: zero-a10g
suggested_hardware: zero-a10g
models:
  - facebook/sam-audio-base
  - facebook/sam-audio-small
  - facebook/sam-audio-large
tags:
  - audio
  - audio-separation
  - sam-audio
  - audio-processing
  - audio-restoration
  - deep-learning
  - pytorch
short_description: AI-powered audio separation using text descriptions
---

# �� Audio Restoration V3

**AI-Powered Audio Separation using SAM-Audio**

## 🎯 What is Audio Restoration V3?

This app uses **SAM-Audio (Segment Anything Model for Audio)** from Meta AI Research to separate specific sounds from audio files based on natural language descriptions.

Simply:
1. 📤 Upload an audio file
2. ✍️ Describe the sound you want to extract (e.g., "singing voice", "drums", "background noise")
3. 🎵 Get two outputs: the extracted sound (Ghost) and everything else (Clean)

## ✨ Features

- 🎯 **Text-Guided** - Separate sounds using natural language
- 🚀 **Multiple Models** - Small (fast) / Base (balanced) / Large (quality)
- �� **Memory Optimized** - Lite model reduces VRAM from 11GB to ~4GB
- 📊 **Long Audio Support** - Chunked processing for extended audio files
- ⚡ **ZeroGPU** - Dynamic GPU allocation for cost efficiency

## 🖥️ Example Descriptions

- `"singing voice"` - Extract vocals from a song
- `"drums"` - Isolate drum tracks
- `"background noise"` - Separate ambient sounds
- `"human speech"` - Extract speech from a recording
- `"guitar"` - Isolate guitar from a mix

## 🙏 Credits

- **SAM-Audio**: [Meta AI Research](https://github.com/facebookresearch/sam-audio)
- **Gradio**: [Gradio Team](https://gradio.app)
- **ZeroGPU**: [Hugging Face](https://huggingface.co/docs/hub/spaces-zerogpu)