Ruach Studio 发布首个候选版本：围绕 YuE2 的本地歌曲工作室

We spent the last weeks building a studio around YuE2, the open song model by m-a-p, and today it reaches its first release candidate. Write the style and the lyrics, and it composes, sings and renders the song on your own card. Nothing leaves your machine unless you point the Writer at a cloud chat model.

What it does

• Whole songs, up to 8 minutes. On one RTX 3090: a 6:12 song in 126 s, and a 7:25 song with its score written first in 198 s.
• The score first, and yours. YuE2 writes the melody and chords as ABC before a note sounds. You can edit them, transpose them, or bring your own score or MIDI.
• Two seeds. Keep the song (the music seed) and hear it rendered anew (the sound seed).
• LoRA training in the studio, on your own songs, unquantized (bf16), with telemetry that tells you which epochs to hear first. Adapters stack on measured roads, under a measured ceiling.
• A guard against garbage. A broken score is caught in seconds and the run is stopped before the GPU is spent on it, and you are told why.
• Post-production, all local: spectrum, artifacts, debuzz, stems (BS-Roformer, htdemucs), remaster, upscale (UniverSR), and a lyrics check by Whisper. One chain runs them all.
• Into your DAW (experimental): a REAPER project with the stems, the score as MIDI, the tempo, the sections as regions and the lyrics on the timeline; DAWproject for Waveform and Bitwig.
• A Librarian for every take; a Writer with versions and a chat model (local or OpenRouter); a cheat-sheet of 200 instruments probed by ear; the API and an MCP server; the page in 7 languages.

What it is not (yet). It is less polished than SUNO out of the box: - a mix can buzz (Debuzz helps); - lyrics can drift (the lyrics check finds where); - some instruments YuE2 plays thinly or not at all (a LoRA teaches them).

You need an NVIDIA GPU (24 GB for everything at full precision), Linux, and about 120 GB of disk for the models, LoRAs and workspaces.

Licences. The code is under AGPL-3.0-or-later. The YuE2 weights are CC BY-NC 4.0, and that licence speaks of the weights, not of the songs made with them: read it before you sell.

What comes next (rc2 and after)

• The Artist room: covers in three shapes at once from one seed, the title and the artist written on them.
• Five more languages for the page: Chinese, French, Portuguese, German, Japanese; right-to-left ones later.
• Voice adapters trained on spoken voices, named by the kind of voice, on Hugging Face.
• The Writer's models with their prices as you type; calmer rooms (dialogs, tips, one shape for the icon buttons).
• A desktop app: an installable page first, then Electron; native plugins for REAPER, Waveform and Bitwig.

Links - Code: https://github.com/igrbible/Ruach_Studio - Models (pinned, checked): https://huggingface.co/goldhub/Ruach_Studio_Models - Site: https://ruachstudio.igr.bible - The full guide, room by room, is inside the studio and in docs/GUIDE.md.

Built on: - YuE2 by m-a-p; - yue2.cpp by ServeurpersoCom; - YuE2 Kit v12 by IronWolve (the base of the page and the scripts).

Every one of our changes is numbered and documented (HERESY 1001–1167). Issues and PRs are welcome. We would most like to hear how it runs on machines that are not ours.