StemSplit AI

Founded by Maxim Van Roy.

Branding, ownership and export metadata are protected by a central branding.json, CI validation, Electron startup tamper checks and backend export metadata signatures. See docs/BRANDING_PROTECTION.md.

Professional local desktop app for DJs and producers to split vocals and stems for Ableton Live workflows.

Stack

• Frontend: React, Vite, TailwindCSS
• Desktop: Electron
• Backend: Python, FastAPI
• Audio engine: Demucs htdemucs by default, FFmpeg / FFprobe

Prerequisites

• Node.js 20+
• Python 3.10+
• FFmpeg installed and available on PATH
• Demucs installed in the backend Python environment

macOS:

brew install ffmpeg

Backend Python environment:

cd backend
python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
python -m pip install demucs

Windows PowerShell:

cd backend
py -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt
python -m pip install demucs

Install Frontend

cd frontend
npm install

Run Locally

Backend:

cd backend
uvicorn main:app --reload

Frontend:

cd frontend
npm run dev

Electron:

cd frontend
npm run electron

API

• POST /upload uploads MP3/WAV/FLAC/AIFF/M4A and extracts duration, sample rate, channels and waveform data.
• POST /split queues Demucs separation with GPU preference and CPU fallback.
• GET /status/{job_id} returns progress, stems and export links.
• GET /download/{job_id} downloads the exported stems folder as a zip.
• GET /download/{job_id}/{filename} downloads an individual stem.

Export Structure

exports/
  song-name/
    vocals.wav
    drums.wav
    bass.wav
    other.wav
    instrumental.wav

Reserved Feature Hooks

The UI and architecture reserve room for BPM detection, key detection, AI mastering, Rekordbox export, AI mashup assistant and cloud rendering.

Vocal Enhancement

Vocals are post-processed after Demucs with a DJ-focused enhancement chain. The default Studio mode keeps vocals open and natural with light cleanup, subtle de-essing, controlled harshness reduction, warmth/body EQ, presence and air enhancement, gentle compression, stereo cleanup, limiting and loudness normalization. Fast uses a lighter chain, and Ultra HQ adds slightly more cleanup while preserving top-end detail.

Producer Studio

StemSplit AI includes an AI co-producer workspace with BPM estimation, key and Camelot detection, harmonic compatibility guidance, drop markers, arrangement sections, mashup suggestions, vocal placement advice, DJ transition ideas and Ableton metadata export. The Producer Studio can render cinematic festival intros, BPM-synced build-ups, transition FX and vocal chops. Festival intros are rendered as multitrack arrangements with layered pads, drones, arena crowd atmosphere, cinematic voice/narration where supported by the OS, vocal atmospheres, percussion tension, sub energy, fake climax, uplifters, a tension pause and drop impact. The backend writes markers.json, bpm.json, arrangement.json, festival_intro_arrangement.json, fx.json and producer_analysis.json beside exported stems.