Text-to-music models provide new audio generational and transformational capability, though their control is not a precise art. Embracing the aesthetics of `code bending`, glitch musicians` approach to software manipulation and the sounds of noise music, the open source Stable Audio 1.0 model is manipulated to enable radical new audio outputs. The domain of historical late 1980s Detroit techno is used as a constrained arena for generation; a comparative study is then undertaken of changing model parameters across all 745 layers of the model. Adding random noise to layer parameters is more likely to lead to interesting results than zeroing (pruning) connections or randomly reconfiguring parameters throughout the model, though all manipulations had some chance of creating interesting outputs. Further potential code bending projects for text-to-music engines are discussed.
(See document for exact affiliation information.)
https://aes2.org/publications/elibrary-page/?id=23004
(486KB)
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.
