Desktop app for recording vocals over music tracks with automatic alignment, normalization, and mixing.
My wife likes to sing. The workflow I had was: find a karaoke video, rip the audio in Audacity, record her singing over the minus track, then manually align three tracks (plus, minus, and vocal recording) by nudging waveforms until they match — every single time. It works, but it's tedious. VocalForge replaces all of that with a single app: load a song, optionally extract the instrumental automatically, hit record, and get a mixed demo track out the other end.
- Load a song — import any audio file as the "plus" (original with vocals)
- Automatic vocal separation — extract instrumental (minus) track using Demucs (Meta's source separation model), or load your own minus track manually
- Record vocals — play the minus track through speakers/headphones while recording microphone input, with simple Start / Stop / Finish controls
- Track preview — switch between Song, Minus (sep), Vocal (sep), Minus (import), Vocal, and Mix Result tracks for playback; seek by clicking waveforms or dragging the seek slider
- Automatic alignment — three modes: background music cross-correlation, vocal stem matching (chain alignment through Demucs-extracted stems), or no alignment. Constrained search window prevents false matches on repetitive songs.
- Noise reduction — stem-guided spectral gating with configurable high-pass filter (default 80 Hz) to remove mic rumble and room noise
- LUFS normalization & mixing — normalize loudness (ITU-R BS.1770-4) and blend vocal + instrumental at configurable ratios
- Export & preview — save the final mix as a WAV/FLAC file and immediately play it back in-app for quick iteration
┌─────────────────────────────────────────────────────┐
│ PySide6 GUI │
│ ┌──────────┐ ┌──────────┐ ┌───────────────────┐ │
│ │ Import │ │ Record │ │ Mix & Export │ │
│ │ Panel │ │ Panel │ │ Panel │ │
│ └────┬─────┘ └────┬─────┘ └────────┬──────────┘ │
├───────┼──────────────┼─────────────────┼─────────────┤
│ │ Audio Engine (sounddevice) │ │
│ │ ┌─────────┐ ┌────────────┐ │ │
│ │ │Playback │ │ Recording │ │ │
│ │ │Stream │ │ Stream │ │ │
│ │ └─────────┘ └────────────┘ │ │
├───────┼────────────────────────────────┼─────────────┤
│ Processing Pipeline │
│ ┌───────────┐ ┌───────────┐ ┌──────────────────┐ │
│ │ Demucs │ │ Alignment │ │ LUFS Normalize │ │
│ │ Separator │ │ (xcorr) │ │ + Mix (pyloudnorm│) │
│ └───────────┘ └───────────┘ └──────────────────┘ │
└─────────────────────────────────────────────────────┘
| Component | Library | License |
|---|---|---|
| GUI | PySide6 | LGPL v3 |
| Audio playback & recording | sounddevice | MIT |
| Audio file I/O | soundfile | BSD 3-Clause |
| Vocal separation | Demucs (htdemucs_ft) | MIT |
| ML runtime | PyTorch | BSD 3-Clause |
| Track alignment | NumPy (cross-correlation) | BSD |
| LUFS normalization | pyloudnorm | MIT |
- Python 3.10+
- Audio interface with microphone input (tested with Behringer X-Air 16 + SM58)
- ~2 GB disk for Demucs model weights (downloaded on first use)
- GPU optional but recommended for faster source separation
git clone https://github.com/Artemarius/VocalForge.git
cd VocalForge
python -m venv venv
source venv/bin/activate # or venv\Scripts\activate on Windows
pip install .
vocalforgeTo include Demucs vocal separation (requires ~2 GB for PyTorch + model weights):
pip install ".[separation]"- Import — Click "Load Song" and select the original track (plus). Optionally click "Separate" to extract the instrumental (and cache the vocals stem for alignment), or "Load Minus" to use your own backing track.
- Preview — Use the track selector to switch between Song, Minus, and Vocal. Click anywhere on a waveform or drag the seek slider to jump to a position.
- Record — Select your microphone input, click "Start". The minus track plays; sing along. Click "Finish" when done (or "Stop" to discard and retry).
- Mix — Choose an alignment mode (Background music, Vocal matching, or None), adjust the vocal/instrumental balance and target LUFS, then click "Mix & Export".
MIT — see LICENSE.
Note: VocalForge uses pre-trained Demucs models from Meta Research. The Demucs repository is MIT-licensed; however, the licensing of pre-trained model weights was not explicitly clarified separately by Meta before the repository was archived.
Users are responsible for ensuring they have the rights to process any audio content used with this application.