Open Source
Explore the latest AI open-source projects from GitHub and HuggingFace.
Explore the latest AI open-source projects from GitHub and HuggingFace.
Ruach Studio is a browser-based music workstation built around YuE2, the open song-generation model, running through yue2.cpp on the user's own NVIDIA GPU. You write a style and lyrics, and the studio composes, sings and renders a full song locally; the README says songs, words and settings stay on the machine unless the Writer is pointed at a cloud chat model. The studio is organized in six rooms: a Creator for generating songs, a Writer for drafting lyrics with a chat model, a Refiner for post-production, an Artist for cover pictures, a Librarian for organizing takes, and a LoRA Trainer. Generation can plan the score first as ABC notation that you can edit, transpose or replace with your own score or MIDI, and two seeds let you keep a song's music and re-render its sound. The Refiner covers spectrum and artifact checks, debuzz, stem separation (BS-Roformer, htdemucs), remaster, upscaling with UniverSR and a Whisper-based lyrics check. The LoRA Trainer fine-tunes adapters on your own songs in bf16. An experimental export writes a REAPER project, or a DAWproject for Waveform and Bitwig, with stems, score, sections and lyrics. The README reports timings on one RTX 3090: a 6:12 song in 126 seconds and a 7:25 song with the score planned first in 198 seconds. Requirements are Linux, an NVIDIA GPU with CUDA (24 GB for everything at full precision) and roughly 30 GB of disk. The interface is offered in 12 languages. The studio's own code is AGPL-3.0-or-later, while the models keep their own licenses: the YuE2 weights are CC BY-NC 4.0, and the README tells users to check what that means for songs they sell.