An example of my ongoing practical research effort dedicated to exploring musical qualities in working with generative neural nets for audio, conceived both as hybrid instruments and as (semi-) autonomous actors, is Reykjavík Sunburn.
Here, four different neural audio models, trained on my own musical material (a corpus of electronic music conventionally written and produced), and a private voice dataset are used in an improvisational setting inside Pure Data, a visual audio programming environment.
In Reykjavík Sunburn, I perform Latent Jamming, a real-time improvisation practice with neural audio synthesizers that embraces concepts of algorithmic composition and generative music. I act in real-time inside the models’ latent space, steering mood, density, and rhythm by generating, modulating, and organizing synthetic signal data streams that resemble latent embeddings. By doing so, I aim to replace deterministic composition with guided exploration: tweak, listen, stabilize, vary.
In Reykjavík Sunburn two RAVE and two vschaos2 models are being used:
- Black Latents: a RAVE V2 model trained on the Black Plastics series. The dataset includes 3 hours of drum- and percussion-heavy electronic music. The resulting model generates mainly percussive output with rough textures and high grittiness. In Reykjavík Sunburn, this model is used as a leading asset to generate the rhythmic baseline and general percussive structure.
- Nobsparse: a RAVE V2 model trained on a hybrid dataset of Tech House and sonically sparse Drum & Bass (about 4 hours of audio material). The model’s characteristics are clear, sterile, and lightweight sounds, harmonic textures, and an isolated but dominant low end. Depending on the process development during the improvisation session, this model serves as a secondary texture generator but can also replace Black Latents’ role in the framework.
- VSC2_Nobsparse: this vschaos2 model has been trained on the same dataset as the Nobsparse RAVE model. In Reykjavík Sunburn, it is used to generate interchanging pads and noise textures for transitions or simply to enrich the composition-performance with a harmonic layer.
- VSC2_Martha2023: being the only model trained on voice data (1,5 hours), this vschaos2 model adds a layer of rhythmic, pseudo-vocal sound on top of the otherwise “instrumental” generations of the three other models.
Public performances
Reykjavík Sunburn (Take 1 Redux) received recognition at the AI Song Contest 2025 where it was selected to the finalist shortlist and performed at the award show at Melkweg, Amsterdam.
Live performance of Reykjavík Sunburn at the 7th Conference on AI Music and Creativity (AIMC 2026) at Musikinstrumenten-Museum, Berlin