AudioGhost AI 🎵👻

AI-Powered Object-Oriented Audio Separation

Describe the sound you want to extract or remove using natural language. Powered by Meta's SAM-Audio model.

🎬 Demo

Audio Separation

https://github.com/user-attachments/assets/49248e25-0c56-46ab-a821-2de7f7016bb6

Video Upload

https://github.com/user-attachments/assets/6b8c08a8-c84f-4fc3-83ad-5703f474fc1b

Features

• 🎯 Text-Guided Separation - Describe what you want to extract: "vocals", "drums", "a dog barking"
• 🎬 Video Upload Support - Upload videos and extract/remove audio sources (audio extraction only, not vision-based)
• 🚀 Memory Optimized - Lite mode reduces VRAM from ~11GB to ~4GB
• 🎨 Modern UI - Glassmorphism design with waveform visualization
• ⚡ Real-time Progress - Track separation progress in real-time
• 🎛️ Stem Mixer - Preview and compare original, extracted, and residual audio

🗺️ Roadmap

• 🖱️ Visual Prompting - Click on video to select sound sources visually (Integration with SAM 2)

Architecture

┌─────────────────────────────────────────────────┐
│                   Frontend                       │
│             (Next.js + Tailwind v4)             │
└──────────────────────┬──────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────┐
│               Backend API                        │
│            (FastAPI + Python)                    │
└──────────────────────┬──────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────┐
│              Task Queue                          │
│          (Celery + Redis)                        │
└──────────────────────┬──────────────────────────┘
                       │
┌──────────────────────▼──────────────────────────┐
│           SAM Audio Lite                         │
│    (Memory-optimized Meta SAM-Audio)            │
└─────────────────────────────────────────────────┘

Requirements

• Python 3.11+
• CUDA-compatible GPU (4GB+ VRAM for lite mode, 12GB+ for full mode)
• CUDA 12.6 (recommended)
• Node.js 18+ (for frontend)

💡 FFmpeg and Redis are automatically installed by the installer.

🚀 One-Click Installation (Recommended)

First Time Setup

# Run installer (creates Conda env, downloads Redis, installs all dependencies)
install.bat

Daily Usage

# Start all services with one click
start.bat

# Stop all services
stop.bat

---

Manual Setup (Advanced)

1. Start Redis

Redis is automatically downloaded to redis/ folder by install.bat. If you prefer Docker:

docker-compose up -d

2. Create Anaconda Environment

# Create new environment (Python 3.11+ required)
conda create -n audioghost python=3.11 -y

# Activate environment
conda activate audioghost

3. Install PyTorch (CUDA 12.6)

pip install torch==2.9.0+cu126 torchvision==0.24.0+cu126 torchaudio==2.9.0+cu126 --index-url https://download.pytorch.org/whl/cu126 --extra-index-url https://pypi.org/simple

4. Install FFmpeg (required by TorchCodec)

conda install -c conda-forge ffmpeg -y

5. Install SAM Audio

pip install git+https://github.com/facebookresearch/sam-audio.git

6. Install Backend Dependencies

cd backend
pip install -r requirements.txt

7. Install Frontend Dependencies

cd frontend
npm install

8. Start Services

Terminal 1 - Backend API:

cd backend
uvicorn main:app --reload --port 8000

Terminal 2 - Celery Worker:

conda activate audioghost
cd backend
celery -A workers.celery_app worker --loglevel=info --pool=solo

Terminal 3 - Frontend:

cd frontend
npm run dev

9. Open the App

Navigate to http://localhost:3000

10. Connect HuggingFace

1. Click "Connect HuggingFace" button
2. Request access at https://huggingface.co/facebook/sam-audio-large
3. Create Access Token: https://huggingface.co/settings/tokens
4. Paste the token and connect

Usage

1. Upload an audio file (MP3, WAV, FLAC)
2. Describe what you want to extract or remove:
   • "vocals" / "singing voice"
   • "drums" / "percussion"
   • "background music"
   • "a dog barking"
   • "crowd noise"
3. Click Extract or Remove
4. Wait for processing
5. Preview and download the results

Performance Benchmarks

Tested on RTX 4090 with 4:26 audio (11 chunks @ 25s each)

VRAM Usage (Lite Mode)

💡 High Quality Mode (float32): Better separation quality but uses +2-3GB more VRAM. Enable via the "High Quality Mode" toggle in the UI.

Processing Time

💡 First run includes model download and loading. Subsequent runs use cached models.

Memory Optimization Details

AudioGhost uses a "Lite Mode" that removes unused model components:

Total Reduction: Up to 40% less VRAM compared to original SAM-Audio

This is achieved by:

• Disabling video-related features (not needed for audio-only)
• Using predict_spans=False and reranking_candidates=1
• Using bfloat16 precision by default (optional float32 for quality)
• 25-second chunking for long audio files

Project Structure

audioghost-ai/
├── backend/
│   ├── main.py           # FastAPI app
│   ├── api/              # API routes
│   │   ├── auth.py       # HuggingFace auth
│   │   └── separate.py   # Separation endpoints
│   └── workers/
│       ├── celery_app.py # Celery config
│       └── tasks.py      # SAM Audio Lite worker
├── frontend/
│   ├── src/
│   │   ├── app/          # Next.js app
│   │   └── components/   # React components
│   └── package.json
├── sam_audio_lite.py     # Standalone lite version
├── QUICKSTART.md         # Quick setup guide
└── README.md

API Reference

POST /api/separate/

Create a separation task.

Form Data:

• file - Audio file
• description - Text prompt (e.g., "vocals")
• mode - "extract" or "remove"
• model_size - "small", "base", or "large" (default: "base")

Response:

{
  "task_id": "uuid",
  "status": "pending",
  "message": "Task submitted successfully"
}

GET /api/separate/{task_id}/status

Get task status and progress.

GET /api/separate/{task_id}/download/{stem}

Download result audio (ghost, clean, or original).

Troubleshooting

CUDA Out of Memory

• Use model_size: "small" instead of "base" or "large"
• Ensure lite mode is enabled (check for "Optimizing model for low VRAM" in logs)
• Close other GPU applications

TorchCodec DLL Error

• Downgrade to FFmpeg 7.x
• Ensure FFmpeg bin directory is in PATH

HuggingFace 401 Error

• Re-authenticate via the UI
• Check that .hf_token exists in backend/

License

This project is licensed under the MIT License. SAM-Audio is licensed by Meta under a research license.

Credits

• SAM-Audio by Meta AI Research
• Core Optimization Logic: Special thanks to NilanEkanayake for providing the initial code modifications in Issue #24 that made VRAM inference reduction possible.
• Built with ❤️ using Next.js, FastAPI, and Celery