StemDeck: free, local 6-stem separator on Demucs, no uploads
Drop in a file or a YouTube link and split vocals, drums, bass, guitar and piano on your own GPU — an Apache-2.0 alternative to Moises and LALAL.AI.
Copy markdown
Six stems, zero uploads
StemDeck runs entirely on your own machine — no account, no quota, nothing leaves your computer. Feed it MP3, WAV, FLAC, M4A or MP4, or paste a YouTube URL, and it splits the track into vocals, drums, bass, guitar, piano and "other."
It's Demucs under the hood
The engine is Meta's open htdemucs_6s model on PyTorch, so separation quality tracks the familiar Ultimate Vocal Remover results. It auto-detects your hardware — CUDA on NVIDIA, Metal (MPS) on Apple Silicon, CPU everywhere else. First launch pulls a ~500MB Python runtime and the ~170MB model, then it is fully offline.
A DAW-style mixer, not just a splitter
Past separation you get a multitrack mixer — mute, solo, level, waveform zoom and loop regions — and can export individual stems or custom mixes. That turns it into a practice, transcription and remix tool rather than a one-shot extractor.
The free answer to Moises and LALAL.AI
It's Apache-2.0 and aimed squarely at cloud splitters like Moises and LALAL.AI. You give up their mobile apps and last-mile polish for zero cost, zero rate limits and full privacy — unreleased tracks never touch someone else's server.
Part of an offline-stem wave
StemDeck lands alongside other new local separators such as Trama, all built on Demucs/UVR-class models — a sign self-hosted audio AI is now good enough to replace the subscription tools for most bedroom-producer work.