Spaces:
Sleeping
Sleeping
Voice Isolation AI — Project Idea
Concept
Build a voice isolation / denoising web app using Meta's pretrained Demucs model, wrapped in a simple full-stack pipeline. No model training from scratch — focus on building the pipeline and a live demo.
Recommended Stack
| Layer | Tool |
|---|---|
| AI Model | Meta's Demucs (pretrained, open source) |
| Backend | Python + FastAPI |
| Frontend | Basic HTML/JS (file upload + audio playback) |
| Deployment | Hugging Face Spaces (free) |
Build Plan
| Step | What to do | Skill shown on CV |
|---|---|---|
| 1 | Install & run Demucs on audio files via Python | Python, audio processing |
| 2 | Build a FastAPI endpoint that accepts audio uploads | Backend API design |
| 3 | Simple frontend to upload audio & play cleaned result | Full-stack thinking |
| 4 | Deploy on Hugging Face Spaces | Deployment, MLOps basics |
Realistic timeline: 2–3 weeks of evenings.
Key Concepts
WebRTC
Browser-native protocol for real-time audio/video streaming (used by Google Meet, Discord). Useful for a future real-time version of this project.
ONNX Model
A format for exporting trained models (from PyTorch/TensorFlow) so they can run anywhere — browser, mobile, edge — without Python or a GPU server. Good for a future client-side version.
Why This Works for a Beginner ML Dev
- No math-heavy training required
- Engages with a real production-grade ML model
- End result is a live, linkable demo — more impressive than code alone
- Covers Python, APIs, basic frontend, and deployment in one project
Future Enhancements (after basics)
- Real-time processing via WebRTC + ONNX Runtime in the browser (fully client-side)
- Fine-tune Demucs on a specific domain (call centers, hearing aids)
- Add a React frontend for a polished UI
Created: 2026-05-01