Below are 10 RAVE models that the plug in initially came with, as I have mentioned in a previous blog, I didn’t find these models to be particularly exciting for the kinds of sounds I wanted to make. Although, I will say, they still produced very interesting and intriguing noises, built from a diverse set of sources. They were useful in helping me understand both how the plug in works and the parameters in which you have control over.
After realising I wanted to explore further into other models I began looking online, I found Ircam forums to be the best source of finding them. The link below is where I retrieved the ones I used from:
https://forum.ircam.fr/search/?topics=RAVE%20Model%20Challenge%202025
From this page I began looking into the published models that were put forward for the Rave model challenge 2025, from the many to look through, there were a few that sounded most interesting to me. I was attracted to the original source of sounds that the data sets were built off and from this I chose to use these main three:
Schizophrenia.ts – Zhao Jiajing and Bryan Yueshen Wu
“Our RAVE model, Schizophrenia, is trained on a 62-minute corpus consisting of audio extracted from a continuous screen recording of YouTube Shorts, captured under a newly created account. The corpus features an unbiased selection of short video reels in a continuous flow, representing the average soundscape an individual encounters while using such social platforms.”
I actually found the concept of this model to be more exciting than the sounds it produced, although I did like the tone and the noises it was generating, I was attracted more to the model because of its data content. I thought using an amalgamation of audio from youtube shorts was such an interesting notion, very relevant of contemporary times.
The audible businesses of social media has become extremely normalised, and that has happened very quickly, it is something that is incredibly unnatural, our brains are processing audio at a speed it is not used to. Therefore I thought by combining this with AI, it creates this incredibly synthetic concept of work, rich with context, meaning and significance, this is the aspect of the model that inspired me.
Black Latents – Martin Heinze
“Types of sounds used: Full length audio tracks from Black Plastics, a seven-part release series comprising a total of 28 electronic music tracks by the artist Martsman, mapping a sonic landscape oscillating between Experimental Techno, Breakbeats, and Drum & Bass.
Total duration of audio corpus used for training: Approximately 3h
Artistic intention: Extract dominant characteristics from a defined body of musical work and use the trained model to respawn new audio material.”
This model was the winner of the challenge and it was apparent that this was a much more musically trained set, made to sound more rhythmically inclined, despite the fact that these models don’t really have a rhythm to them because of the randomisation and latent noise of AI. It did however have more distinct sounds which acted as a more solid foundation to the noise of the model (compared to the others which feel looser and more flighty).
086-jaap-3.5m-noisefloor – Jaap Blonk
“This RAVE model is trained on a dataset of Jaap Blonkʼs vocal performances. The dataset was recorded by Blonk with Jonathan Reus specifically for training models as part of Reusʼs Dadasets project, in particular for text-to-voice synthesis in Reus and Victor Shepardsonʼs Tungnaá instrument.
The dataset totals about 100 minutes of close-mic recordings of Blonkʼs solo voice performances, which are a mix of scored and improvised.”
I was attracted to this model due to the fact it was trained on vocals, not instruments, I became intrigued into how it would react to other vocals being inputed through the model. I knew I wanted to work with my voice for this piece as I wanted to create this kind of natural vs unnatural element to the work, and so when I found this model, I thought it was fitting to my themes.
At the time of writing this I have not yet experimented with inputting my voice through this model, so I am currently unsure of how it will sound but I am intrigued to see if it will have an interesting effect.