NEBULA: A PCA-BASED METHOD TO EXPLORE RAVE-ENCODED AUDIO REPRESENTATIONS

The challenges of exploring the latent spaces of deeplearning audio models often lead artists to rely on chance, randomness, and combinatorial approaches, making it difficult to steer these models toward musically meaningful outcomes. In this paper, we explore how Principal Component Analysis (PCA) applied to pre-encoded RAVE (Realtime Audio Variational Autoencoder) latent representations can provide a more controlled and curated approach to navigating these high-dimensional spaces. By restricting exploration to selected regions of the latent space, musicians gain clearer pathways to achieving specific sonic goals. Although t-SNE and UMAP effectively preserve intricate local structures, we show how the linearity, computational efficiency, and interpretability of PCA offer distinct advantages for real-time applications. In addition, we introduce a graphical user interface (GUI) and a sensor system for manipulating ’timbral vectors’ derived from PCA components, providing an intuitive tool for identifying, refining, and shaping sonic transformations. To evaluate the effectiveness of PCA, we systematically compare its performance with t-SNE and UMAP, highlighting the tradeoffs among these methods.