Stemyla splits a song into its instruments, in the browser, on your own GPU. Drop an MP3 or WAV, press "Separate 6 stems", and get bass, drums, other, vocals, guitar and piano as separate lanes you can mute, solo, slow down (without changing pitch), loop, and save as WAV or MP3. Nothing leaves the computer.

It is written in Rust and compiled to WebAssembly, including the GPU code: the separation model runs on WebGPU through hand-written WGSL kernels. 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.

Open the published page (for example https://beachape.com/stemyla/) in a browser with WebGPU: Chrome or Edge on macOS, Windows or ChromeOS, Safari 26, or Firefox 141 and later on Windows (status by platform). It has been tested in Chrome on macOS (a Radeon Pro 5500M and an Intel UHD 630); other browsers should work but are untested.

1. Press Download model (350 MB) once. The page keeps the model in the browser's own storage.
2. Drop a song (44.1 kHz) on the page, or press Demo song.
3. Press Separate 6 stems. On a laptop GPU a 3.5-minute song takes about 7 minutes.
4. Use Mute and Solo per lane, the tempo slider, and drag across the bar above the lanes to loop a section. Save on a lane downloads it as WAV or MP3.

Each song you drop, and its stems once separated, stays in the browser, so a reload brings it back. The Storage menu lists the songs kept, opens or forgets one, shows the space used, and can remove the model. About links to the model, its code and its paper.

Page options, added to the address: ?selftest checks the GPU against reference results (when served with the reference files, see below); ?bench times the GPU kernels; ?gpu=integrated asks for the low-power GPU.

The separation model is BS-RoFormer SW, a six-stem BS-RoFormer whose architecture is open source (MIT, lucidrains/BS-RoFormer). The trained weights are a different matter: their trainer is unknown and no licence was ever stated for them (see this open question). So this repository and the published page do not contain them. "Download model" fetches a copy that already exists on Hugging Face (lumabeat/bs-roformer-sw) and checks its SHA-256 before use; "Load model file…" accepts a local .safetensors file instead. Whether you may use the weights is for you to judge.

A PyTorch checkpoint (BS-Rofo-SW-Fixed.ckpt) converts to the file the app reads with:

cargo run -p xtask --release -- roformer-convert BS-Rofo-SW-Fixed.ckpt bs_roformer_sw.f16.safetensors

Everything is built and served with Rust tools.

rustup target add wasm32-unknown-unknown
cargo install --locked trunk                                   # --locked matters: without it the build of Trunk itself fails
(cd app && trunk build --release)                              # into app/dist
cargo run -p xtask --release -- serve app/dist --port 8080     # http://localhost:8080/

To use ?selftest, make the reference files once and serve them too:

cargo run -p xtask --release -- roformer-selftest bs_roformer_sw.f16.safetensors third_party/roformer/selftest
cargo run -p xtask --release -- serve app/dist --port 8080 --mount /selftest=third_party/roformer/selftest

The native app (cargo run -p app --release) plays songs and stems but cannot separate; separation is web only.

The page is static, so a gh-pages branch serves it. One command builds it for the site's path, commits the build to gh-pages and pushes that branch:

cargo run -p xtask --release -- publish        # for https://<user>.github.io/stemyla/ or <custom domain>/stemyla/

It refuses to run with uncommitted changes, so every published version names the commit it was built from. It commits through a temporary git index, so the checkout you work in is never switched or touched, and it builds on the remote's gh-pages if another clone published last. Options: --public-url /other-path/ for a different path, --no-push to commit locally only, --remote <name>, --branch <name>, --allow-dirty.

The first time, choose the gh-pages branch under the repository's Settings → Pages. A project site on an account whose user site has a custom domain (beachape.com here) appears under that domain at /stemyla/. The site is about 6 MB; the model is not part of it.

cargo test --workspace                       # unit, integration and UI snapshot tests (Linux needs ALSA headers)
cargo clippy --workspace --all-targets

• roformer and roformer-gpu compare the network with PyTorch when the fixtures exist (tools/roformer_export.py, a Python script that only generates test data; tests skip without them).
• The GPU kernel tests run on any Vulkan, Metal or DX12 GPU. Without one, SwiftShader's Vulkan driver works: VK_ICD_FILENAMES=<path>/vk_swiftshader_icd.json WGPU_BACKEND=vulkan cargo test -p roformer-gpu --release.
• tools/*.mjs check the built page in headless Chromium (Node and Playwright): web_selftest.mjs, worker_check.mjs, web_session.mjs, web_audio_resume.mjs, web_bench.mjs, web_separate.mjs, web_smoke.mjs.

MIT or Apache-2.0, at your option (LICENSE-MIT, LICENSE-APACHE). Third-party crates and their licences are listed in THIRD-PARTY.md. The model weights are not covered; see "The model".