StemDeck: A Free, Local, Open-Source Alternative to Cloud AI Stem Splitters

StemDeck is an open-source desktop app that breaks a song into as many as six isolated tracks—vocals, drums, bass, guitar, piano, and everything else—entirely on the user’s own hardware. It accepts common audio and video formats or a pasted YouTube link, then drops the results into a DAW-style multitrack mixer with per-stem volume, mute, solo, looping, waveform zoom, and live VU meters. The separation itself runs on Meta’s open-source Demucs (htdemucs_6s) model, auto-selecting CUDA, Apple Silicon MPS, or CPU depending on the machine, with an optional lead/backing vocal split via a UVR-MDX-NET model.

The pitch is squarely a privacy and cost play against cloud services like Moises and LALAL.AI: no account, no upload, no quota, no subscription. Because processing happens locally, audio never leaves the device, and the project explicitly states it does not store, cache, or redistribute downloaded content. Beyond separation, it layers on song analysis—BPM detection via librosa, key and scale estimation, and BS.1770 loudness measurement. The stack is Python 3.12 with a FastAPI backend, a vanilla-JS/Web Audio frontend with no build step, and Tauri v2 desktop shells for macOS and Windows.

The maintainer is candid about scope: StemDeck covers the core local use case and deliberately stops there, pointing users who need higher quality, mobile access, or deeper musician tooling toward commercial products. The YouTube-import feature is the main legal caveat, framed as a convenience for content users already have the right to process rather than a downloader. Pre-built installers ship with each GitHub Release, downloading the ~170 MB model on first launch, with source builds available for Linux and GPU setups.