Stemyla

Stemyla does one thing that used to require a studio: it splits a song into its individual instruments — bass, drums, other, vocals, guitar and piano — entirely in your browser, on your own GPU. Drop an MP3 or WAV onto the page, press "Separate 6 stems", and you get six lanes you can mute, solo, slow down without changing pitch, loop a section of, and save as WAV or MP3. Nothing ever leaves your computer: the page, the model and your songs all stay local.

The engineering is the interesting part. The app is written in Rust and compiled to WebAssembly — including the GPU code. The separation model is BS-RoFormer SW, running on WebGPU through hand-written WGSL kernels, which is how it manages serious neural audio separation with no server involved. On a laptop GPU, separating a 3.5-minute song takes about seven minutes. The page needs a one-time 350 MB model download (kept in the browser's own storage), and it runs in Chrome or Edge on macOS, Windows and ChromeOS, Safari 26, or Firefox 141 and later on Windows.

The backstory is pure vibe coding. The author's own repository description tags it "Vibe-coded!", and the README explains it plainly: "It started as a winter-holiday project for learning to build software with AI, and to hear the guitar part of a song on its own." The repo went up on October 13, 2026, and the published page is live with a demo song you can separate in one click.

One honest caveat the author flags himself: the trained model weights come from an unknown trainer on Hugging Face with no stated license, so the repo ships the app without them and leaves it to you to judge whether you may use the weights. For learners, remixers and the merely curious, that transparency is part of the charm — an AI-built tool for pulling music apart, with its rough edges documented in the open.