Murzinograph: Navigating Sound through Latent Space Visualizations
Gustavo Guzmán
- poster
- Type: medium
- Session: Poster Session 3
Abstract:
We present the Murzinograph, a proof‑of‑concept system that maps audio spectrograms into human-interpretable, three‑dimensional visualizations through a bespoke convolutional autoencoder. By constraining latent space dimensionality, the Murzinograph privileges contour‑like structures over exact spectral reconstruction, exposing a �Visualization/Reconstruction trade‑off’ whereby relaxed reconstruction fidelity yields more semantically useful manifolds. We examine across musical and bioacoustic datasets, and present a community case study that demonstrates how collective interpretation of visualised acoustic data can nurture sound and noise awareness. Furthermore, we discuss implications for Interactive Machine Learning and eXplainable AI in NIME contexts, and outline avenues toward hybrid, conditioned and generative model scalings that balance interpretability and synthesis. We conclude by illustrating how this work is positioned within these fields as an accessible tool for creative endeavours, particularly in the context of the Global South.