Examensarbete (master) Stockholm : Kungliga Tekniska högskolan, 2025

"The Benjolin, a chaotic synthesizer, presents a challenge due to its non-intuitive parameter-to-sound mapping. This thesis investigates a timbral similarity-based control interface for the instrument. Timbral and loudness descriptors, extracted from a corpus of 85,536 Benjolin recordings using a bag-of-frames approach, were used to train a variational autoencoder (VAE), generating a 3D latent space, which was visualized as a navigable point cloud. This was compared to a UMAP-generated baseline. A qualitative user study with 19 participants, analyzed with reflexive thematic analysis, revealed that the timbre-based control interface exhibited high learnability and explorability, but comparisons between the VAE and UMAP clouds were inconclusive. Furthermore, users experienced difficulties navigating the 3D space, and they expressed a need for greater explainability to fully utilize the interface. An inspection of the latent spaces showed that the chosen audio descriptors had failed to fully capture the Benjolin’s chaotic behavior, limiting timbral similarity within the point cloud." - abstract