rave-scope-plugin

IRCAM RAVE realtime audio VAE nodes for Daydream Scope.

RAVE (Realtime Audio Variational autoEncoder, Caillon & Esling, IRCAM) is a neural audio model that encodes waveforms into a compact latent space and decodes them back. Trained checkpoints cover speech, percussion, orchestral instruments, field recordings, and more; manipulating the latent between encode and decode gives real-time timbre transfer and neural audio effects.

This plugin exposes RAVE to Scope's node graph so you can build patches like:

AudioSource → RAVE Timbre → Sink

or the fully split variant for more control:

AudioSource → RAVE Encode → RAVE Latent Manip → RAVE Decode → Sink
                                  ▲
                             curve / OSC / MIDI

Installation

Install as an editable Scope plugin (development):

uv pip install --editable C:/_dev/projects/rave-scope-plugin

Or install from a git URL in Scope's plugin UI:

git+https://github.com/<you>/rave-scope-plugin

Restart Scope after install. The RAVE nodes appear under the rave category.

Getting RAVE models

RAVE models are TorchScript .ts files exported from acids-ircam/RAVE with --streaming. The plugin scans:

~/.daydream-scope/rave-models/*.ts

(overridable via DAYDREAM_SCOPE_RAVE_MODELS_DIR). Drop .ts files into that directory and they show up in the RAVE Loader dropdown.

Sources of pretrained checkpoints (all CC-BY-NC-4.0, non-commercial):

• IRCAM Forum — https://acids-ircam.github.io/rave_models_download
  • Direct download: https://play.forum.ircam.fr/rave-vst-api/get_model/<name>
  • Names include: vintage, percussion, VCTK, darbouka_onnx, nasa, sol_ordinario, musicnet
• HuggingFace — Intelligent-Instruments-Lab/rave-models — https://huggingface.co/Intelligent-Instruments-Lab/rave-models
  • birds_dawnchorus, voice_multivoice, guitar_iil, organ_bach, water_pondbrain, many more

Example download from the IRCAM forum (curl is fine):

mkdir -p ~/.daydream-scope/rave-models
curl -L "https://play.forum.ircam.fr/rave-vst-api/get_model/vintage" \
     -o ~/.daydream-scope/rave-models/vintage.ts

Nodes

Node	Purpose
RAVE Loader	Load a .ts model; outputs a rave handle
RAVE Timbre	One-shot audio → latent-manip → audio (demo headline)
RAVE Encode	Audio → RAVE latent
RAVE Decode	RAVE latent → audio
RAVE Latent Manip	Per-dim bias/scale + additive noise on a latent
RAVE Prior	Autonomous generation from the model's prior (--prior exports)

All audio-producing nodes output at 48 kHz stereo to match Scope's audio bus. The plugin internally resamples to the model's native rate (usually 44.1 or 48 kHz), runs inference in the model's native channel count, and lifts the result back to stereo.

RAVE Timbre — per-dim controls

RAVE Timbre and RAVE Latent Manip expose eight per-dim bias_* and scale_* sliders plus a global noise and wet control. RAVE's first ~8 latent dimensions carry the perceptually salient axes (brightness, attack, roughness, vowel colour, etc. — axes vary per model). Sweep them live and the timbre morphs.

Credits & licensing

This plugin is MIT-licensed. The RAVE code and pretrained models are licensed separately under CC-BY-NC-4.0 (non-commercial use). Do not redistribute the .ts files commercially; users must download their own.

Original RAVE work: Caillon & Esling 2021, arXiv:2111.05011. See github.com/acids-ircam/RAVE and the ACIDS-IRCAM team's publications for background.