Repository navigation
Ruach Studio 2.0.9
Released on 19 November 2026 · the release · CHANGELOG · the visiting card · the announcement
Since 2.0.8 the Studio has taken more than 120 numbered changes (HERESY 2605 to 2855). Most of them you will not see; they are the reason a room opens faster or a take is not lost. A few change what the Studio is. This is the walk through them, room by room. The full list is in CHANGELOG.md, and each change, with what was measured, in heresy/docs/HERESY.md.
Until now the Studio lived on Linux, and on Windows only through WSL2. WSL2 asked too much of a Windows machine: virtualisation switched on, a distribution registered, and a Studio inside it that a browser outside could not always reach. So we stopped asking and built the Studio for Windows itself.
Windows 11 gets one zip. Unpack it into a short folder, double-click RuachStudio.cmd, and the first start brings Python, the Studio's packages and the models a song needs into that folder, nowhere else. The zip carries two engines: CUDA 12.8 for NVIDIA cards from Pascal (the GTX 10 series, compute 6.1) to Blackwell, and one for the processor. The start looks at what nvidia-smi sees and takes the right one. On a Pascal or Volta card the songs sing on the card while the lab's own work (splitting stems, writing a score from a recording) runs on the processor, because PyTorch's CUDA wheels start at compute 7.5. A Windows user whose name is written in Cyrillic was our first test case, and the Studio works there.
A Mac with Apple Silicon gets one zip too: the Metal engine for macOS 13.3 and later and RuachStudio.command. The model's quant follows the Mac's memory. It is not yet tried on a real Mac: GitHub's Mac runner checks the package (every binary arm64, the lab answers), but no one has sung a song with it on a real Apple Silicon Mac yet. If you have one, tell us what you see.
Both are a beta. They are made and tried for the Musician and the Librarian: write, sing, listen, compare, keep. The rooms that lean on Linux services (the Trainer's runs, the Narrator's engines, the Studio's own writer model, the Artist and the Videographer) follow one by one. Neither package is signed, so Windows SmartScreen and macOS Gatekeeper will ask once; the release notes say which button to press. We would rather ship you an honest beta now than a polished promise later.
The AppImage for Linux also got a fix you may never notice unless you run Arch, CachyOS or Fedora's newest: it used to take two libraries from the system, and a system that builds them for a newer glibc stopped it at load. Now it carries every library it needs; only the NVIDIA driver is your system's. Thanks to SanFranRan for issue #10.
Vocal over My Backing. Bring an instrumental of your own into Cover or Remix. The Studio reads its key, tempo, bars, chords and parts (a guide vocal on it is split off first), YuE2 writes its own melody over those chords and sings your lyrics, and the voice is laid onto your backing bar by bar. You get a new take with both stems for your DAW. Your backing itself is never touched. The melody is the model's; the Studio says under the buttons what it cannot promise.
The song's history. The Versions button opens every version of a title that was sung, what changed between them and their takes. Two versions compare line by line, and any of them goes back into the form.
Re-sing from the break. The song's passport finds the rest in the voice nearest a moment (3:00 by default) and regenerates from there, everything before it kept.
Several scores before any sound. Plan two to four scores at once, each on its own seed, and see them side by side: bars, sections, breaks, which section sings. The longest whole one is offered to the form.
Re-denoise and Re-decode. A take keeps its music codes and its latents. Make its sound again through another solver, or through another decoder in seconds.
And the style got a ceiling. YuE2 is not a model you feed a page of adjectives: 200 characters without spaces is the ideal, over 400 we do not endorse, and past 800 nothing sings until the style is shortened. Nothing is ever cut for you; the counter just turns red.
The Studio's player used a limiter to hold peaks when you raised the volume past 100 % or lifted a take by a few dB. On loud passages that limiter pumped the low end: the sub-bass came and went. A take sounded different in the Studio than in any other player. The players now carry no compressor, no limiter and no Web Audio at all. At a gain of 1 what you hear is the file; with +dB or the volume past 100 % it is the take raised by that plain gain and clipped at full scale, as the engine's old peak clip cut it, played by the browser's own audio element. The compare window's wave now grows and shrinks with the lift each version is heard at.
Your own voice is now cut on its wave: drop a clip, drag the 30 seconds you want, hear them, keep only that. Higgs v3 holds its voice through a long narration, each piece read with the one before joined to its reference. Greek goes to the engine polytonic, as written. Drama 3 can read a narration in whole paragraphs and a dialog's turns together, up to three voices a pass. A piece that stalls or loses its end is read again, and the take's note says so.
Its curves together keep the scale of their own charts, values show under the cursor again, a run started by hand is shown and steered from the room, and the run list answers in a blink while a run trains. A score-mode run whose tracks have no sheet now stops and says so instead of quietly teaching something else, an idea from leeQzz's fork. The Studio is also ready for a grown 4.6B backbone beside the 3B ones; ours is still in training and is not part of this release.
We audited what the page does while you are not looking. It was doing a lot. Now a tab you left asks for the cards once in 50 seconds instead of every two, the bars and canvases redraw only what moved, and the Library answers from what did not change. An idle room went from 1,346 layouts a minute to 74, and the Library's catalog from 6.0 s to 0.3 s in the lab.
To leeQzz, whose fork gave two changes that are now ours too: the lyrics' guidance passed through the public API, and the Trainer's guard for runs without a sheet. Forks that come back as ideas are the best thing open source does. To manchuwook for issue #9 with its root cause, to SanFranRan for issue #10, to Timon for the fixes in his yue2.cpp and the Windows build recipe, and to everyone who tried the first native betas and told us what they saw.
- Linux:
git pullor Pinokio's Update; Engine → Updates offers 2.0.9. Or the new AppImage. - Windows 11 and macOS: the zips on the release page.
Sing one short song on Windows or a Mac and tell us here your card, your system and the time it took. That is how a beta stops being one.
Questions and findings about this release are welcome in its announcement, in Discussions and in Issues.
Release articles
Articles
Mini-guides
Links