Skip to content

Ruach Studio 2.0.0

Choose a tag to compare

View all tags
@igrbible igrbible released this 25 Oct 21:05
· 8 commits to main since this release
v2.0.0
ba7fec4

The first stable release. Ruach Studio's own fine-tune is the default model; songs are edited as the model writes them; the Refiner works as a canvas, and the DAW and the piano roll come in as a beta; there is a playlist, the Trainer runs on several cards, and every string is in all twelve languages (HERESY 1567–1797, after 2.0.0-pre).

The model

  • Our own fine-tune is the default: YuE2-RS-FW, the First Wave, is our training of m-a-p's YuE2-3B on a dataset we assembled (32,633 tracks, about 1,800 hours, a score for almost every track). It is published as merged weights at seven checkpoint and strength pairs in goldhub/YuE2-RS-Finetunes, each as GGUF (BF16, Q8_0, Q6_K) and as BF16 safetensors. 2.4 × 0.75 comes with a fresh install at the size your card takes (BF16, Q8_0 or Q6_K; a Q5_K_M install keeps m-a-p's alone). The others are in Engine › Models and LoRAs › More models. The card on the Hub holds the measurements and will name the best merge after the blind listening. The adapters themselves are not published.
  • Settings you already have stay yours. A studio that already names a backbone in its settings keeps it; m-a-p's weights and the IRA and IRAE merges stay where they are.

The Creator

  • The piano roll (beta):
    • the score twice on paper above two weighted keyboards;
    • chords with Shift, the computer's keys and MIDI;
    • real instruments: the Salamander Grand for the instruments, a choir for the voice that sings the vowels and the lyrics' syllables;
    • Transpose a segment.
  • Songs edited as they are written:
    • Only the opening (your bars, the model plans on from them);
    • Extend… (a take goes on past its end);
    • Re-sing this line, a section sung once more, Sing a window anew;
    • a take arranged anew from its sections;
    • Chords (Off · Gentle · Varied · Bold);
    • the chord bed;
    • the voice transposer.
  • Versions beside the Title, one tree per song; Compare up to twelve takes, rated and noted.
  • Sound decode, experimental: a decode ladder in Sampling that renders the sound in overlapping pieces. It sounds different from the stock decode, which stays the default everywhere else.
  • AB2 at 48 steps as the default solver, and Full sings an octave lower by default.

The Refiner and the DAW (beta)

  • The Refiner is a canvas of nodes: your own way through it, pipelines, a target node, clones, profiles, eight stems.
  • New nodes: Voice upscale (UniverSR, Apollo-vocal, FlashSR), Voice tamer, Voice attack and body, Bass under the kick, Details up, Mono Maker, Peak glue and a true-peak ceiling.
  • The DAW after a sound engineer's pass:
    • buses, VCA groups, sidechain, automation;
    • inserts and Print, parallel compression, a strip's gate, returns with pre-delay;
    • values typed into the console;
    • LRA and PLR on the master;
    • the song's sections on the ruler.
  • Every download carries its tags; Content Credentials (C2PA) sign them.

The Librarian and the page

  • The playlist with a crossfade; Regenerate keeps a take's own backbone.
  • The workspace column scrolls on its own, with its headings in the chapter look.
  • The page opens only when everything has loaded.
  • MIDI for the whole studio: a keyboard plays the piano roll, a controller's DAW port runs the transport, and plugins may bring scripts (early support).

The Trainer

  • Several cards at once (DDP), every save a moving average (EMA), saves in tenths of an epoch, Pause and Resume.
  • Ranks for each half, tracks that teach the sound half only, a stop at the plateau, then the sound half alone.
  • Datasets as parquet.

Engine

  • yue2.cpp merged with upstream (8af0174) with our patches kept whole: the same sound, bit for bit, at the same speed.
  • Intel Arc as a beta (SYCL), shown in Engine → GPUs.

Install / update

  • New: INSTALL.md, or Pinokio on Linux (download it from https://github.com/igrbible/Ruach_Studio and press Install). The install takes about 33 GB (60 GB with the Artist's painters). The default fine-tune adds 3.8 GB on a 16 GB card and 7.2 GB on a 24 GB one.
  • From 2.0.0-pre or an rc: Engine → Updates in the page, or Pinokio's Update. The models you have stay. Engine › Models and LoRAs fetches the fine-tune: the «fw» part, or one merge with heresy/fetch-heresy.sh --fw 2.4x0.75-Q8_0.

Linux with an NVIDIA card of 8 GB and more (6 GB with shorter songs: see INSTALL.md); Intel Arc as a beta. Windows 11 through WSL2: INSTALL_WINDOWS.md.

Known limits

  • The DAW and the piano roll are a beta: a project opened in REAPER or Bitwig is not yet tested by hand, and when a take's stems load while the DAW's drawer is shut, the lanes do not shrink for the console over the whole window until the console is shown again (2.0.x).
  • Sound decode's ladder is experimental, and its output differs from the stock decode.
Assets 4
Loading