🎵 MusicAI

Desktop app for musicians — stem splitting, BPM & key detection, audio file manager.

Built with Python (Flask + pywebview) and vanilla JS.

Features

• Stem Splitter — split any track into vocals/instrumental (2-stem) or drums/bass/other/vocals (4-stem) using Demucs htdemucs
• Mixer view — after splitting, each stem gets its own waveform with Mute, Solo and volume controls
• BPM detection — multi-method voting (Essentia RhythmExtractor + PercivalBpmEstimator + librosa), analyzes middle 60% of track to avoid intro/outro noise
• Key detection — multi-profile voting across 4 Essentia profiles (temperley, krumhansl, edma, bgate) + harmonic signal boost; falls back to librosa chroma if Essentia unavailable
• Camelot wheel — click the key badge to toggle between note name and Camelot code (e.g. Am минор ↔ 8A)
• Analysis cache — results stored in .analysis_cache.json, invalidated on file change; re-analyzes if key confidence < 75%
• File manager — browse, play, rename, delete audio files; quick access to Desktop/Downloads/Music
• Bottom player bar — shows BPM and key for currently playing track

Requirements

• Python 3.9+
• macOS or Windows 10/11

Installation

git clone <repo>
cd musicai

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate

pip install torch demucs librosa soundfile tqdm flask pywebview

Optional (macOS, improves BPM/key accuracy significantly)

pip install essentia

On Windows, essentia is not available via PyPI — the app falls back to librosa automatically.

Run

python app.py

First run downloads the htdemucs model (~300MB).

Supported formats

MP3 · WAV · FLAC · AIFF · M4A · OGG · OPUS

Notes

• Processing runs on CPU by default; GPU speeds it up significantly
• Analysis cache is saved to .analysis_cache.json in the project root (gitignored)
• On Windows, pywebview uses EdgeWebView2 (built into Windows 10/11, no extra install needed)

License

MIT