A free, open-source tool called StemDeck has been drawing attention on Hacker News this weekend, climbing the front page with nearly 200 upvotes and passing 3,000 stars on GitHub. Its pitch is simple: professional-grade AI stem separation that runs entirely on your own computer, with no account, no uploads, and no subscription.
Stem separation — pulling a song apart into isolated vocals, drums, bass, and other instruments — has become a standard part of modern music workflows. DJs use it for live remixing, producers use it for sampling, musicians use it for practice and transcription. Until now, the most polished options have been cloud services like Moises and LALAL.AI, which process your audio on their servers for a monthly fee. Pour plus de contexte sur cette histoire, consultez notre tendances IA.
StemDeck takes the opposite approach, and its README states it plainly: "Free, local stem separation. No account. No upload. No subscription."
How StemDeck Works
Under the hood, StemDeck uses Demucs, Meta Research's open-source music separation model — specifically the htdemucs_6s configuration, which splits audio into six stems: vocals, drums, bass, guitar, piano, and other. The application automatically detects the best available hardware, using CUDA acceleration on NVIDIA GPUs, Apple's Metal Performance Shaders on Apple Silicon Macs, and falling back to CPU when neither is present.
Getting started requires little more than dragging a file onto the import bar. StemDeck accepts MP3, WAV, FLAC, OGG/Opus, MP4, and M4A files, and can also import from a YouTube URL — a convenience the project limits explicitly to content users have the right to process.
Once separated, the stems open in a DAW-style multitrack interface. Users can mute and solo tracks, balance levels, zoom into waveforms, loop regions, and export individual stems or a custom mix. The waveform editor renders min/max samples across all stems with shared normalization and stem-aligned lanes, features usually found in paid digital audio workstations rather than free utilities.
The tool also supports stem subset extraction: clicking stem chips selects which stems to keep, and clicking from a fully selected state snaps to just one stem, with further clicks adding or removing tracks from the selection. It is a small touch, but it turns the exporter into a genuinely flexible tool for building custom backing tracks or practice mixes.
Built for Practice, Remixing, and Study
The project positions itself for musicians, producers, and hobbyists who need isolated parts of a track for practice, transcription, remixing, and creative audio workflows. A bass player can mute everything except the drums to practice with a rhythm section; a producer can pull an acapella for a remix; a teacher can isolate a melody for a transcription exercise.
Because the separation models run on whatever hardware is available — from gaming PCs with NVIDIA cards to Apple Silicon laptops — the same project covers bedroom producers and professional studios alike, without the per-minute processing limits that cloud services impose.
Why Running Locally Matters
The local-first design is more than a cost decision. When audio never leaves your machine, there is nothing to upload, no processing quota to hit, and no server storing your files. For working musicians handling unreleased material, that privacy guarantee is the feature.
"Everything happens locally and nothing leaves your machine," the project's documentation emphasizes. The project also draws a clear line on what it is not: StemDeck describes itself as a stem separation tool, not a downloader, and says it does not store, cache, or redistribute any content processed through its YouTube import feature.
That positioning has not kept the project small. StemDeck's repository shows more than 3,000 GitHub stars since it was created in May 2026, and the project maintains an active community across Discord, Reddit, and X.
A Free Alternative in a Subscription Market
StemDeck arrives as an explicit alternative to the subscription model. The project compares itself directly to cloud stem-splitters like Moises and LALAL.AI, acknowledging that commercial products still win on polish, mobile apps, and deeper musician tooling — while StemDeck covers users who want stems for personal study without an account or recurring fee.
Fait inhabituel pour un projet de ce niveau d'attention, StemDeck déclare qu'il n'accepte aucun argent, parrainage ou financement de qui que ce soit. Il n'y a pas de niveau payant à vendre et aucune entreprise derrière l'outil - juste une base de code open source, une licence permissive et une communauté d'utilisateurs.
Fait partie d'une tendance plus large de l'IA locale
StemDeck s'inscrit dans un modèle plus large d'outils d'IA : les modèles performants sont devenus suffisamment petits et efficaces pour fonctionner sur du matériel grand public, et les développeurs les regroupent dans des outils qui troquent la commodité du cloud contre la confidentialité et le contrôle des coûts. La génération d'images locales, la transcription locale et les modèles de langue locale ont tous suivi cet arc, et la séparation audio est un choix naturel : Demucs est ouvertement disponible depuis des années, en attendant que quelqu'un construise le flux de travail autour de lui.
Pour les musiciens, les producteurs et les amateurs, le calcul est facile à comprendre. Si vous possédez déjà un ordinateur raisonnablement performant, StemDeck le transforme en un studio de séparation de tiges qui ne coûte rien et ne rapporte rien à personne.
Le projet est disponible sur GitHub, avec des téléchargements pour les principales plates-formes et de la documentation sur son site Web à l'adresse stemdeck.app.
---
Restez à la Pointe de l'IALes dernières actualités, analyses et percées en IA — au même endroit.
Lire plus d'actualités IA →
