Firefox FIXED! This was due to Firefox's legacy handling of certain browser features and networking protocols, which triggered false positives and connection errors during development.
While the site is fully functional on Firefox, its architecture is more prone to these networking "false flags" compared to modern Chromium-based browsers. If you experience connection issues, please check your local network settings, otherwise, the tool works as intended! (there is no future support for firefox browsers as more modern chromium-based browsers are preferred here)
A professional, 100% free, web-based tool that separates audio tracks into individual stems (Vocals, Drums, Bass, Other) using the state-of-the-art Demucs AI engine.
This project was built to provide a high-quality alternative to paywalled services like Lala.ai or Splitter.ai, running entirely on volunteer hardware with no file limits.
🔗 Try it now: https://vicsanity623.github.io
- 🚫 No Paywalls & No Limits: Upload long tracks (FLAC, WAV, MP3) without "pay-per-minute" restrictions.
- 💎 Dual AI Models:
- ⚡ Speed Mode: Uses standard
htdemucsfor fast results (~2 mins). - 💎 Ultra Quality: Uses
htdemucs_ft(Fine-Tuned) for audiophile-grade separation with minimized bleed.
- ⚡ Speed Mode: Uses standard
- 📱 PWA Ready: Installable as a native app on iOS and Android.
- 🌊 Interactive Player: Real-time waveform visualization using WaveSurfer.js with "Solo Mode" playback.
- 🔒 Privacy First: Files are processed in RAM on a secure backend and deleted immediately after download generation.
This is a headless implementation of Meta's Demucs, orchestrated via a custom Python backend and served securely over the public internet.
- Frontend: Hosted on GitHub Pages (Static HTML/JS).
- Tunneling: Uses Tailscale Funnel to create an encrypted pipeline from the user to the local server.
- Backend: A Python Flask API running locally on an Intel iMac.
- Queue System: Implements a FIFO (First-In-First-Out) queue to manage multiple users on a single GPU/CPU resource.
| Mode | Model ID | Description |
|---|---|---|
| Speed Mode | htdemucs |
Hybrid Transformer. Great balance of speed and quality. Best for sketching ideas. |
| Ultra Quality | htdemucs_ft |
Fine-Tuned. Heavier neural network trained on a larger dataset. Focuses on cleaner high-end frequencies and vocal isolation. |
Please Read Carefully
This service runs on personal hardware, not a cloud farm.
- Queueing: If multiple users upload simultaneously, you will see a "Waiting in queue" message. Please be patient.
- Processing Time:
- Speed Mode: ~2–3 minutes per song.
- Ultra Mode: ~5–8 minutes per song (due to heavy computation).
- Availability: If the site hangs or fails to connect, the host machine may be offline for maintenance.
This tool is intended for educational, research, forensic, and production use on content you own or have permission to modify.
- ✅ You must own the rights to uploaded audio.
- ❌ Do not upload copyrighted music without explicit permission.
- ✅ You are fully responsible for the usage of the separated stems.
Legal Notice We do not store user content. All files are transient and wiped after processing. Using this tool to infringe on copyright is strictly prohibited.
- Frontend: HTML5, CSS3, JavaScript
- Visualization: WaveSurfer.js
- Backend API: Python Flask
- Secure Tunnel: Tailscale Funnel
- AI Engine: Demucs (Meta Research)
This project relies on the incredible work by the Meta Research team:
@article{defossez2021hybrid,
title={Hybrid Spectrogram and Waveform Source Separation},
author={Défossez, Alexandre},
journal={arXiv preprint arXiv:2111.03600},
year={2021}
}