Project 05 — Desktop · AI Audio

OctoSplitter

A desktop app that splits a song into its stems.

OctoSplitter logo: a glossy red octopus with equalizer bars
PythonDesktopAudio MLWindows + macOS

About the project

OctoSplitter (formerly StemSplitter) is a desktop app for Windows and macOS that takes a song (MP3, WAV, FLAC, M4A, AAC, OGG, Opus, AIFF, WMA) and splits it into one file per stem: vocals, drums, bass, other, plus guitar, piano, instrumental and a no-drums mix.

It uses BS-RoFormer SW to split the song in one pass and, in Maximum quality, averages in a second vocal model, MelBand-RoFormer. Vocals, drums, bass and other add back up to the original song exactly.

Goal

Make clean stem separation easy for anyone: drag in songs, pick the stems and quality, and get the files, with no command line and no cloud upload.

What it does

State-of-the-art separation

Three presets (Balanced, Maximum, Fast). On the MUSDB18 test set Balanced averages 10.20 dB SDR, against 9.14 dB for the previous version.

Uses your GPU

NVIDIA CUDA on Windows and the Metal GPU on Apple Silicon, with a CPU fallback and a memory governor that keeps the system responsive.

Batch and playback

Queue songs or folders, play each stem with mute, solo and volume, and export MP3 (320, 256 or 192 kbps) or 24-bit WAV.

Updates itself

Checks GitHub releases, downloads only changed files, verifies the new version and rolls back if the check fails. Available in five languages, plus a CLI.

Tech stack

Application

Python
PySide6 (Qt)
PyTorch
RoFormer models
ffmpeg

Packaging & delivery

PyInstaller
Inno Setup (Windows installer)
macOS app bundle
GitHub Actions
GitHub Releases

Quality

Regression tests
Ruff
Self-test used by CI

Screenshots

← Back to projects