DESCRIPTION:
RAVE (Realtime Audio Variational autoEncoder) is an algorithm designed for real-time, high-quality audio waveform synthesis using neural networks. It leverages a variational autoencoder (VAE) architecture, which compresses audio data into a compact latent representation, allowing efficient reconstruction of audio signals.
Key features of RAVE include:
- Fast, high-quality audio generation: It excels at producing accurate audio in real-time, making it ideal for interactive applications (20x real-time at 48 kHz sampling rate on standard CPU)
- Real-time use: Integrated with tools like Max and Pure Data (Pd), RAVE can be used with the nn~ decoder for real-time sound generation and transformation. A VST plugin makes it easy to use in any DAW.
- Applications: Common uses include audio synthesis, timbre transformation, and style transfer.
In short, RAVE is a powerful tool for real-time audio generation, offering both speed and quality.
In just a few months, RAVE popularized the creation of models based on audio recordings, thanks in particular to the publication of a series of tutorials and open-source code. A growing and ebullient community of users took hold of the algorithm, and numerous models emerged. Although these models can be quite costly to produce (around twenty GPU hours), very few have so far been published, often due to copyright issues. This challenge concerns models trained on personal recordings for which the authors own all rights.
The aim of this challenge is to support the authors of the best models and to collectively establish a repertoire of RAVE models, enabling everyone to benefit from the richness and variety of approaches in the field of timbre/music transfer.
The challenge is hosted by the DAFNE+ platform, which promotes content using NFTs.
A public vote awards three prizes to participants.
PRIZE:
The awards ceremony will take place during the IRCAM Forum Workshops 2025, between March 26 and 28, 2025 at IRCAM, Paris.
- 1st award: 2000€ plus one year IRCAM Forum Premium Membership
- 2nd award: 1000€ plus one year IRCAM Forum Premium Membership
- 3rd award: 500€ plus one year IRCAM Forum Premium Membership
If multiple entries receive the same number of winning votes, their prizes and the following ones will be shared among them. For example:
- If two candidates tie for the highest score and a third has the next highest, the first two will share (2000+1000)/2 = €1500 each, and the third will receive the third prize of €500.
- If one candidate has the most votes (€2000 first prize) and three candidates tie for the second-highest votes, their prize will be (1000+500)/3 = €500 each.
IMPORTANT DATES:
- Call publication in November 2024 on forum.ircam.fr and on dafneplus.eu
- DAFNE+ Upload Platform opened from the 1st of December 2025 (noon CET) to January 31, 2025 (midday CET) February 10, 2025 (noon CET) - Deadline extension
- Public vote from the 11th of February 2025 (noon CET) to the 28th of February 2025 (noon CET).
- Award ceremony in March 2025
SUBMISSION:
To participate, participants must upload an application to the content manager of the DAFNE+ platform, with the following content in a single zip file, with the “AI model” type:
- The model in .ts format. Mode “forward” only.
- Model description: A description of the model in term of
- Types of sounds used (free description, instruments, genre, playlist...)
- Total duration of audio corpus used for training
- Artistic intention: Do you want to achieve something special with this corpus?
- A picture, artwork, photo presenting the model.
- Free additional information
- Model training copyright: A letter of intent specifying respect for copyright under CC BY-NC license (see below) and declaring third-party sources if used.
- Model examples: A set of output audio files showing the effect of the model:
- 5 free 15sec generations, in MSprior or decoder mode
- 5 transformations in forward mode of 5 imposed sounds, downloadable via the following links:
- singing twinkle twinkle, Mr. moon.wav by bectec -- https://freesound.org/s/665123/ -- License: Creative Commons 0
- 106 BPM Drum Loop 1.wav by esares -- https://freesound.org/s/431874/ -- License: Creative Commons 0
- intertwined 0T_50mm by Setuniman -- https://freesound.org/s/165172/ -- License: Attribution NonCommercial 4.0
- deep house drum beat.wav by djfroyd -- https://freesound.org/s/349708/ -- License: Attribution 3.0
- 15-Second Strum by ViraMiller -- https://freesound.org/s/745885/ -- License: Attribution 4.0
- Short Biography (400 words max., in English) and High-Definition photo of the author.
To submit your model to the challenge on the DAFNE+ platform, please follow this tutorial.
A submission template is available in the competition content.
Only complete proposals will be considered.
EVALUATION
The three prizes will be awarded by vote of members registered on the DAFNE+ platform (free registration), rewarding the three models with the highest number of votes (in descending order for the 3 prizes). The models will be published on the DAFNE+ platform Marketplace with tag “RAVE Model Challenge”. From February 1, 2025, members will be able to download the models to evaluate them, as well as listen to the audio files to vote for their favorite model. The link to the voting platform will be provided on February 1, 2025 and voting will close on February 28 (noon CET).
LICENSING CONDITIONS OF THE SUBMITTED MODELS
The RAVE models submitted for the competition will be published with free access (no bitcoin fee) on the DAFNE+ platform under the Creative Commons V4 license with option BY-NC.