Skip to content

Ruach Studio 2.0.9

Latest

Choose a tag to compare

@igrbible igrbible released this 19 Nov 09:47
Immutable release. Only release title and notes can be modified.

2.0.9 takes the Studio off Linux for the first time: native packages for Windows 11 and for a Mac with Apple Silicon, as a beta beside the Linux Studio, with no WSL2 and no Docker. The AppImage now runs on the newest Linux libraries (Arch, CachyOS, Fedora; issue #10). The Musician sings a voice over an instrumental of your own, keeps a song's history, re-sings from the break nearest a moment, plans several scores before a sound and makes a take's sound again from what it kept. The Narrator cuts your own voice on its wave and keeps Higgs v3's voice through a long narration. The Trainer draws its curves together on their own scales. The page rests when you do. And the players play a take as it is, with no compressor or limiter in the way (HERESY 2605–2800 and the fixes after them).

New

  • The Studio on Windows 11, natively (beta) (HERESY 2649, 2685, 2686): one zip; unpack it into a short folder and double-click RuachStudio.cmd. The first start brings Python 3.12, the lab's packages and the models a song needs into that folder only; every start after it opens the Studio in your browser. Two engines ride in the zip: CUDA 12.8 for NVIDIA cards from Pascal (compute 6.1) to Blackwell, and one for the processor (AVX2); the start picks by what nvidia-smi sees. A user name or a folder in Cyrillic works (the engine reads its paths as UTF-8 on Windows).
  • The Studio on a Mac with Apple Silicon (beta) (HERESY 2687): one zip with the Metal engine for macOS 13.3 and later and RuachStudio.command; the model's quant is chosen by the Mac's memory (Q8_0, Q6_K or Q5_K_M).
  • Vocal over My Backing (HERESY 2677–2681): in the Musician's Cover or Remix, bring an instrumental of your own (a file, or a take of the Library or its instrumental stem). The Studio reads its key, tempo, bars, chords and parts (a guide vocal on it is split off first), the model writes its own melody over those chords and bars and sings your lyrics, and the voice is stretched bar by bar onto your backing and mixed over it as a new take, both stems kept for your DAW. Your backing itself is never changed.
  • The song's history (HERESY 2689): the Composer's Versions button opens every version of a title that was sung, what changed between them, and their takes; any version back into the form, or two compared line by line.
  • Re-sing from the break near a moment (HERESY 2690): 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 (HERESY 2691): 2 to 4 scores planned at once on their own seeds and shown side by side (bars, sections, breaks, which section sings); the longest whole one is offered to the form with its seed.
  • Re-denoise and Re-decode (HERESY 2692): a take's sound made again from what it kept, through another solver and steps, or through another decoder in seconds.
  • Your own voice in the Narrator, cut on its wave (HERESY 2624, 2643): drop a clip, see its wave, drag up to 30 seconds of it, hear it, and only that piece is kept.
  • Every sound brought in passes one gate, by its bytes (HERESY 2659): WAV, FLAC, OGG, MP3 and phones' voice notes (M4A, AAC, 3GP, AMR) up to 200 MB, each told by its header, never its name, searched for a hidden program or archive and decoded whole before it is kept; anything else is refused with the reason.
  • A take's datasheet says its licence, part by part (HERESY 2658): its model, its decoder, each LoRA, a Narrator's engine, and M-A-P's permission for creators where a song is concerned.

The Musician

  • The lyrics' density against their genre (HERESY 2693): each line's genre ceiling beside the meter, a gauge per section, the song's beats per second. They only show; nothing is changed for you.
  • A style's ceiling of 800 characters without spaces (HERESY 2799, 2800): the counter beside the label is green to 200, amber to 400, red to 800; past 800 nothing is cut, but nothing sings from the field until it is shortened. 200 is the ideal; YuE2 is not a model you feed a page of style.
  • The engine's queue holds 12 songs by default, up to 20; Generate while songs run asks first (add, stop the running one, or cancel); «Stop all Musician tasks» under the lamp by the logo (HERESY 2674–2676).
  • Extend… goes on past a take that one window held whole, in pieces, to the length you ask (HERESY 2666). Multistep (AB2) is the sound solver everywhere, the engine's default too; Euler is labelled the draft (HERESY 2670). The denoising steps default to 48 in the engine as in the page (HERESY 2663).
  • A failed computation stops the song instead of making silence (HERESY 2626, after Timon's yue2.cpp). A backbone whose file disappears no longer takes the engine down (HERESY 2662).
  • The Takes column paints 50 takes at a time, so «From all Workspaces» opens in a fraction of a second on a library of thousands (HERESY 2671).
  • The compare window plays each version from memory and takes 24 at once (HERESY 2627, 2629).
  • The Studio is ready for a grown 4.6B backbone: a YuE2-4.6B-*.gguf in the models is offered beside the 3B ones and the engine switches between them (HERESY 2660). Ours is still in training and is not part of this release.

The Narrator

  • Higgs v3 holds its voice through a long narration: each piece is read with the one before joined to its reference (HERESY 2695).
  • Greek is read as written: polytonic, its breathings and accents kept (HERESY 2696); the guide says where Drama 3 misses a stress (HERESY 2697).
  • Drama 3's Window and Scene boxes: a narration read in whole paragraphs in a 16,384-row window, a dialog's turns read together up to three voices a pass (HERESY 2698, 2699).
  • A piece that stalls (over 4 s of silence) or loses its end is read again, and the take's note says so (HERESY 2700, 2701).
  • Tags suggested by the Writer's model: a ✨ button puts the engine's own tags ([sighs] where someone sighs) into the text, taken only when no word changed, Ctrl+Z to take it back (HERESY 2694).
  • An MP3 of a spoken take sounds as its WAV again (HERESY 2605).

The Trainer

  • The curves together keep the scale of their own charts, a box per curve to hide or light it, LR and memory in a chart of their own (HERESY 2657, 2664, 2665); the charts say their values under the cursor again (HERESY 2653).
  • A run started by hand in its own unit is shown as running, with its step, time left, Pause and Stop (HERESY 2654); the run list answers in a blink while a run trains, and the Trainer opens on the run that is training (HERESY 2655, 2656).
  • A Full or Melody run whose tracks have not one sheet stops and says so, instead of teaching Direct alone under the score mode's name (HERESY 2762; the idea from leeQzz's fork).

The page rests when you do

  • The page knows when you are away (the tab hidden, or five idle minutes with nothing playing) and slows or stops its repeating work, coming back the moment you look (HERESY 2721). A tab at rest asks for the cards once in 50 s, not every 2 s; no sample of the graphs is lost (HERESY 2727).
  • The system bar, the jobs' strips, the spectrum's playhead and the Refiner's canvas redraw only what moved (an idle room: 1,346 → 74 layouts a minute; the idle Refiner 13.0 → 7.8 s of the browser's minute) (HERESY 2722–2725, 2728).
  • The Library opens without reading every take again (the catalog 6.0 → 0.3 s), the covers' list is lighter and asked once at a time, and the Trainer loads its runs in a cascade (HERESY 2669, 2726, 2729–2732, 2734).

Around the rooms

  • The players play the take as it is (HERESY 2835–2838): no compressor, no limiter and no Web Audio in any player of the Studio, so the low end no longer pumps on loud passages. At a gain of 1 you hear the file itself; +dB and the volume past 100 % are a plain gain, the take raised 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 grows and shrinks with the lift each version is heard at (HERESY 2838).
  • The score guard (HERESY 2830–2834): the music's floor never runs past the end of its score, a plan with no bar holding a note is said as a missing key line is, and the inspector finds near silence and shows the wave past the score.
  • Every prompt that writes a style keeps the style's length rule, and the Writer asks a model back once for a style it wrote too long (HERESY 2840, 2841); negative words in the style are marked while you type and asked about before a song, and the guide says how not to prompt (HERESY 2845, 2847).
  • The live spectrum behind the player is drawn from the take itself (HERESY 2849); the decode ladder never runs past its window (HERESY 2851).
  • One field height, 34 px, in every room (HERESY 2793); the Studio's own checkboxes and radios, drawn by the theme (HERESY 2688); the Librarian's numbers on one line, unfolding on hover (HERESY 2792).
  • The Library's eye hides and shows the hidden and the disliked, and remembers (HERESY 2647).
  • The Guide opens our own documents inside itself (HERESY 2648), and its Hebrew reads right to left on its own in every language (HERESY 2791).
  • One way to name a processor everywhere: GPU0 · RTX 3090, XPU0 · Arc Pro B50, and the backend by its own name (HERESY 2733).
  • The audio upscale works again (HERESY 2646, 2646b; found with its root cause by manchuwook in issue #9).
  • The public API passes the lyrics' own guidance (cfg_lyrics) on to the engine (HERESY 2761; the idea from leeQzz's fork).
  • From Timon's yue2.cpp, with thanks (MIT): the music cache in Q8_0 at about half the memory, off by default (HERESY 2633); tinyBLAS on by default (HERESY 2634); a GPU self-test at start that falls back to the processor and says why (HERESY 2635).
  • Every sampling field shows its bounds (HERESY 2636); the transcriber reads in F32 by default, Q8_0 a choice (HERESY 2625); a Cover from a recording that is too loud is levelled first, and says so (HERESY 2651).

The AppImage

  • It runs on Arch, CachyOS and Fedora's newest libraries (HERESY 2798, issue #10): it used to take zlib and libcrypt from the system, and a system that builds them for a newer glibc stopped the engine and the lab at load with «GLIBC_ABI_DT_RELR not found». The image now carries every library its programs need; only the NVIDIA driver is the system's, and a system folder on LD_LIBRARY_PATH no longer comes before the image's own.
  • Its own window is light unless your desktop is set to dark (HERESY 2616), as the AppImage catalog asked.

Windows and Mac: what works, and the limits of a beta

  • Made and tried for the Musician and the Librarian: write, sing, listen, compare, keep. The first Windows beta zip ran end to end in a Windows 11 virtual machine, and a tester runs it on a Quadro P5200. 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) come one by one in the next betas. A chat server of your own or an OpenRouter key can still write for the Writer.
  • Older NVIDIA cards (Pascal, Volta: GTX 10xx, Quadro P, Titan V): songs sing on the card, but the lab's PyTorch builds for compute 7.5 and up only, so the lab's own work (splits, transcription) runs on the processor, more slowly.
  • Unsigned: neither package is signed. On Windows, SmartScreen asks once: More info → Run anyway. On a Mac, Gatekeeper stops the first start: right-click RuachStudio.command → Open, or System Settings → Privacy & Security → Open Anyway.
  • Paths: unpack on Windows into a short folder (for example C:\RuachStudio): Windows' 260-character path limit reaches into the Studio's Python.
  • The Mac: not yet tried on a real Mac. The package is checked on GitHub's Mac runner (every binary arm64, macOS' own libraries only, the lab answers), but no song on a real Apple Silicon Mac has been reported yet. Tell us what you see.
  • No updater yet in the native packages: a new beta is a new zip.

What to try

  • Bring an instrumental into Cover → Vocal over My Backing and let the Studio sing your lyrics over it.
  • Open a song's Versions and compare two of them line by line.
  • Plan four scores at once and pick the longest whole one.
  • Take a song you like and Re-decode it with RS x3 v2 or Standard, in seconds.
  • Leave the Studio's tab for five minutes and watch your browser's CPU fall.
  • On Windows or a Mac: sing one short song and tell us your card, your system and the time it took.

Install or update

  • Linux, from git or Pinokio: git pull (or Pinokio's Update) brings an installed Studio to 2.0.9, and Engine → Updates offers it.
  • Linux, the AppImage: download RuachStudio-2.0.9-core-x86_64.AppImage with its .sha256, sha256sum -c it, chmod +x it and start it in place of the old one; your models, songs and settings stay where they are, and its first start asks only for what changed. The .zsync brings only the changed blocks.
  • Windows 11: download the Windows zip and SHA256SUMS.txt, check the sum (certutil -hashfile FILE SHA256), unpack into a short folder, double-click RuachStudio.cmd.
  • macOS 13.3 or later, Apple Silicon: download the Mac zip, check its sum (shasum -a 256 FILE), unpack it, open RuachStudio.command as above.

Known limits

  • Windows and the Mac are a beta: some rooms are not there yet (above), and nothing is signed.
  • Linux needs an NVIDIA card with the driver 570 or newer for full speed; an Intel Arc runs the engine as a beta, and with no card the Studio computes on the processor, slowly.
  • Not on free cloud GPUs (Colab, Kaggle): it is a whole studio, not a notebook.

Downloads

  • The source tarball and the tag v2.0.9.
  • RuachStudio-2.0.9-core-x86_64.AppImage with its .sha256 and .zsync.
  • The Windows 11 zip and the Mac zip (beta), with SHA256SUMS.txt.

Thanks

To leeQzz, whose fork gave the lyrics' guidance in the API and the Trainer's guard for runs without a sheet (HERESY 2761, 2762); to manchuwook (Steven Altsman) for issue #9 with its root cause and the question of issue #2; to SanFranRan for issue #10; to Timon (timoncool) for yue2.cpp's fixes and the Windows build recipe; to everyone who tried the first native betas and told us what they saw; to m-a-p for YuE2, ServeurpersoCom for yue2.cpp and ggml-org for ggml; to Fish Audio for Drama 3 (Built with Fish Audio) and to Boson AI for Higgs TTS 3 (Built with Higgs TTS 3 licensed from Boson AI USA, Inc.).

Everything, entry by entry: CHANGELOG.md and heresy/docs/HERESY.md.