StemDeck is a free, open-source stem separator with a DAW-style mixer that runs locally

What StemDeck does

StemDeck is a new free and open-source desktop application that separates audio into individual stems — vocals, drums, bass, guitar, piano and a catch-all "other" channel — using an AI model that runs entirely on the user's own hardware. The project drew attention on the Hacker News front page, linking to its GitHub repository.

According to the project's README, you can drop in MP3, WAV, FLAC, OGG/Opus, MP4 or M4A files, or paste a YouTube link, and the app splits the audio into as many as six stems. Playback happens in a multitrack mixer modeled on a digital audio workstation: stems can be muted, soloed and level-balanced, the waveform view supports zooming and looping over a region, and individual stems or a custom blend can be exported.

The project positions itself as a local, no-cost alternative to cloud services such as Moises and LALAL.AI. It requires no sign-up, imposes no usage quota, and never uploads audio — nothing leaves the machine. Internet access is needed only for the initial model download, roughly 170 MB which is cached afterwards, and for fetching audio from YouTube, which the developers frame as a convenience for content you have the right to process.

Mixer and analysis features

Beyond basic separation, the app supports subset extraction: clicking stem chips chooses which stems to keep, and when you pick a subset, a seventh lane automatically holds the complement — the full song minus the selected stems — for A/B comparison. A summed mix of just the selected stems can be exported as a single WAV.

Each stem gets a volume fader, mute, solo and a monitor option, with live VU meters showing post-gain RMS plus peak hold. The app also analyses songs: tempo via librosa's beat tracker, key and scale with confidence estimates using Albrecht–Shanahan profiles, and loudness measured as integrated LUFS per ITU-R BS.1770 alongside sample peak in dBFS. Jobs can be cancelled mid-pipeline with partial files cleaned up, and a library panel organises tracks by folder with drag-and-drop, search and trash.

How it's built

The separation engine is Demucs htdemucs_6s, Meta AI's open-source six-stem neural network, with automatic device selection across CUDA on NVIDIA GPUs, MPS on Apple Silicon and CPU as a fallback. The backend is Python 3.12 managed with uv, serving REST and Server-Sent Events through FastAPI, while the desktop shells are built with Tauri v2 — Rust with WKWebView on macOS and WebView2 on Windows. YouTube fetching runs through yt-dlp, transcoding and mixing use FFmpeg, and an optional lead/backing vocal split uses the UVR-MDX-NET Karaoke 2 model from the Ultimate Vocal Remover project.

Pre-built installers cover macOS, with an Apple Silicon build using MPS acceleration on M1 and later plus a CPU-only Intel version, and Windows, which ships as either a roughly 700 MB CPU-only package or a 1.6 GB NVIDIA CUDA variant. First launch downloads a bundled Python runtime of about 500 MB along with FFmpeg and the model; subsequent launches skip setup.

The trade-offs, stated plainly

Unusually, the README includes a candid comparison table against commercial rivals, and it concedes several points. Cloud services generally offer higher-quality proprietary models, run fast regardless of your hardware, support up to ten stems depending on the plan, and add extras such as chord detection, lyrics and mobile apps. StemDeck processes one job at a time with no batch mode, lacks features like pitch shifting, and its own interface is described as hobby-grade rather than production-polished. The pitch is deliberately narrower: private, free stem extraction for practice, transcription and remixing. The project also states that it does not store, cache or redistribute downloaded content, and that it accepts no money, sponsorship or funding.

Why it matters

Stem separation has largely consolidated around subscription cloud services, which by design means sending recordings to third-party servers. StemDeck bundles a proven open model into an approachable desktop app with a genuine mixer, at zero cost and with no data leaving the machine — useful for musicians practising parts, transcribers, and anyone working with audio they would rather not upload. It also illustrates a broader shift: capable AI audio models are now open enough that a well-designed wrapper can compete on convenience and privacy even while conceding on raw quality. For producers on Apple Silicon or with NVIDIA GPUs, it is a practical, no-commitment tool worth keeping installed.