Release of RAVE Models Repository

IIL releases trained RAVE neural synthesis models from various projects.
Wed Oct 18 2023
The RAVE logo in IIL colors

The RAVE logo in IIL colors

Check out the Hugging Face repo here!

RAVE (which stands for Realtime Audio Variational autoEncoder) is a popular neural synthesis model designed by Antoine Caillon and the ACIDS research group at IRCAM, Paris. When you train a RAVE model on some audio data, it learns an encoder which extracts compressed features from the audio, and also a decoder which takes them back to sound. You can either use the decoder by itself as a unique kind of synthesizer, or run new audio through the encoder-decoder pair, transforming it to sound more like the training data.

What’s great about RAVE is that it can run in real time on a laptop and is well supported in software like Max/MSP, Pure Data and SuperCollider. We have also developed and released our own SuperCollider and TidalCycles integrations. However, training new RAVE models takes some familiarity with machine learning workflows and also a fair amount of GPU time. Over the past 18 months, we’ve been training RAVE models for projects like the Magnetic Scores, Pluma, Living Looper, Tölvera, and now we’re releasing them to the public.

IIL postdoc Jack Armitage live coding speech RAVE models at Algorithmic Art Assembly 2022.

Hugging Face is a hub for open-source and community machine learning projects – sort of a Github for ML. We decided to put our RAVE models in a Hugging Face repo for discoverability – it’s an obvious place to search for models – and it provides a nice user interface and storage while letting us manage the repo via git.

Our new Hugging Face repo contains models of guitar, saxophone, church organ, magnetic resonator piano, speaking and singing voices, water, birds, whales, and we’ll keep adding more. All our RAVE models inherit the Creative Commons Noncommercial license required by the ACIDS software, meaning you can do just about anything with them except sell the models themselves as part of a product.

The repo also includes some model checkpoints which can be used with our fork of the RAVE code to speed up training via transfer learning.

MIT Media Lab's Jessica Shand & Manuel Cherep perform MOTHERBIRD with IIL postdoc Jack Armitage at TENOR 2023, using flute and bird RAVE models.

Nicola Privato performing Mouija, a piece using Stacco and Thales, two magnetic instruments developed at IIL. The video features magnets, water, speech, choir, organ RAVE models

If use our models, let us know! And remember to cite using:
@misc {intelligent_instruments_lab_2023,
	author       = { {Intelligent Instruments Lab} },
	title        = { rave-models (Revision ad15daf) },
	year         = 2023,
	url          = { },
	doi          = { 10.57967/hf/1235 },
	publisher    = { Hugging Face }

Happy synthesizing!