I Forgot The Words To This Song

I'm making up my own.

A web app for people who forget song lyrics and want to sing their own version instead. Upload a song, remove the vocals, and sing whatever you want over the instrumentals.

Live Deployments

• Backend API: https://huggingface.co/spaces/neonwatty/forgot-the-words-api
• Frontend: (TBD - Static Space)

Current Status (Jan 2026)

The backend is deployed on HuggingFace ZeroGPU but experiencing CUDA compatibility issues with the SAM Audio model. The facebook/sam-audio-small model may have dependencies on internal PyTorch APIs not available in standard releases.

Known Issues:

• RuntimeError: NVML_SUCCESS == r INTERNAL ASSERT FAILED during inference
• ZeroGPU quota limits (daily GPU time)

Potential Solutions to Explore:

1. Switch to Demucs (Meta's production-ready vocal separator)
2. Use a dedicated GPU Space instead of ZeroGPU
3. Wait for SAM Audio to support public PyTorch releases

How it works

1. Upload a song you can never remember the words to
2. AI removes the vocals, leaving just the instrumentals
3. Sing your own lyrics over the music

Tech Stack

• Frontend: React + Tailwind CSS (hosted on Hugging Face Static Space)
• Backend: Gradio + Meta SAM Audio (hosted on Hugging Face ZeroGPU)
• Model: facebook/sam-audio-small

Project Structure

├── backend/              # Original backend code
│   ├── app.py
│   ├── requirements.txt
│   └── README.md
├── hf-backend-temp/      # Current deployed backend (synced to HF Space)
│   └── app.py
├── src/                  # React frontend
│   ├── App.jsx
│   └── index.css
├── demo/                 # Demo files
│   └── rickroll.mp3      # Test audio
├── dist/                 # Built frontend (deploy this)
├── test_api.py           # API test script
└── README.md

Deployment

Prerequisites

• Hugging Face account (Pro recommended for ZeroGPU access)
• Node.js 18+
• HF_TOKEN environment variable set

Backend (HuggingFace Space)

The backend is deployed at: neonwatty/forgot-the-words-api

To update:

# The hf-backend-temp/app.py is the source of truth
# Push changes using the HuggingFace Hub API or git

Frontend

# Update the API URL
echo "VITE_API_URL=https://neonwatty-forgot-the-words-api.hf.space" > .env

# Build
npm install
npm run build

# Upload dist/ contents to a Static Space

Local Development

npm install
npm run dev

Testing the API

export HF_TOKEN='your_token_here'
python test_api.py

GPU Usage

With Hugging Face Pro ($9/month):

• Daily ZeroGPU quota
• ~30-60 seconds per song (when working)

---

Your lyrics are valid too.