<aside> 📍
You are not lost.
You have arrived at a sub-page in the AIxDESIGN Archive.
This is a public repository - our way of working in the open and sharing as we go.
Have fun!
</aside>
<aside>
INDEX
</aside>
A recap of our workshop with Moisés Horta Valenzuela (Hexorcismos), held online on 10 June 2026.
<aside>
Open Culture Tech is an initiative by Thunderboom Records and the Netherlands Institute for Sound & Vision that makes immersive technology more accessible to artists through residencies, open-source tools, and educational showcases. As part of OCT 2.0, AIxDESIGN partnered to host a series of workshops exploring how AI shows up in music production - both its creative potential and the trickier questions around agency, authorship, and how artists actually want to work with these tools.
</aside>
Moisés Horta Valenzuela, aka Hexorcismos, is an artist, technologist, and musician from Tijuana, based in Berlin. He's been making electronic music for almost twenty years and working at the intersection of machine learning and sound for nearly eight. He's the creator of SEMILLA.AI, a neural audio synthesis platform, and his work has been featured at Ars Electronica, MUTEK, and the Royal Academy of Music in Stockholm. His practice sits at the intersection of computer music, AI, and digital technology through a decolonial lens. He’s also the Creative Technologist in our Saz-e-yaad project.
On 10 June 2026, we hosted Moisés online for the most technically dense session in our OCT 2.0 series: a two-hour walkthrough of CoDiCodec-Flow, an open-source pipeline for training and performing with your own generative audio model. The pitch: teaching a model your sonic vocabulary, then jamming with it live.
This was a more technical session than the others - command line comfort assumed, Python familiarity helpful, headphones required, and ideally 20-60 minutes of audio you'd want to train on. The full event page is on Lu.ma. Some things broke along the way (Whereby kept dropping the slide share, Colab GPU runtimes wobbled), but we carried on.
Before any training code, Moisés started with the conceptual foundation. A Variational Autoencoder (VAE) compresses an audio signal into a low-dimensional latent code, then reconstructs it back. Think MP3 compression, except the rules aren't handcrafted by engineers but learned by a neural network. A neural audio codec is what you call a VAE applied to audio.
The interesting move is what happens after training: you can throw away the encoder and use the decoder alone as a synthesiser. Random coordinates in latent space → new sound. Not memorised, not retrieved from a database, but sound that lives in the distribution the model has learned.
"You can think about it as a sort of terrain where the peak of the mountain is the most precise reconstruction. And then you have the valleys where there's very little overlap between probability distributions. So basically this is what makes the space navigable. It's a continuous space. So you can do interpolation, you can do sampling, you can apply arithmetic operations within this latent space representation." — Moisés Horta Valenzuela
https://www.instagram.com/reel/DZ5XYwSoaoG/
To make this less abstract, Moisés demoed Trajectory, an unreleased tool he's building that visualises latent spaces in 3D (using PCA to flatten the 64-dimensional originals into something we can see). He encoded the same two sounds with two different codecs side by side: Music2Latent kept them on clearly separated trajectories; Stable Audio 3 had them overlapping. The point landed without argument - a well-organised latent space takes work off the generative model that has to learn inside it.
CoDiCodec is a neural audio codec from Sony CSL, pre-trained on ~20,000 hours of music. The dataset is internal and undisclosed (presumably Sony catalogue, since they own the copyrights). What's unusual: it produces both continuous embeddings and discrete tokens from a single trained model. Most architectures pick one path. CoDiCodec-Flow uses the continuous path.
A question came up in the chat: how do you feel about using a model trained on undisclosed Sony data? Moisés's answer was technical, and it pointed at something specific about how these architectures actually work:
"To go back to reconstructing the data it was trained on, you would need to guess what these latent coordinates are. We're talking about a 64-dimensional, even 512-dimensional vector. You would need to match every coordinate at the same time, and that's statistically impossible. They're representing the features of the data. They're not a lookup table. They're not a database." — Moisés Horta Valenzuela