Conventional sonifications directly assign different aspects of data to auditory features and the results are not always “musical” as they do not adhere to a recognizable structure, genre, style, etc. Our system tackles this problem by learning orthogonal features in the latent space of a given musical corpus and using those features to create derivative compositions. We propose using a Singular Autoencoder (SAE) algorithm that identifies the most important Principal Components (PCs) in the latent space. As a proof-of-concept, we created sonifications of ionizing radiation measurements obtained from the Safecast project. Although the system successfully generates new compositions by manipulating the latent space, with each principal component changing different musical aspects, these changes may not be readily noticeable by listeners, despite the PCs being mathematically decorrelated. This finding suggests that higher-level features (such as associated emotion, etc.) may be needed for better results.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=22256
(254KB)
Click to purchase paper as a non-member or login as an AES member. If your company or school subscribes to the E-Library then switch to the institutional version. If you are not an AES member Join the AES. If you need to check your member status, login to the Member Portal.
