Stable Ecologies?: Computational Audible Infrastructures in Generative AI Audio

Generative AI tools and models, such as Stability AI’s audio (Evans, 2024) or OpenAI’s Jukebox (Dhariwal, 2020), are prevalent in audio culture. Owned by platforms, these systems enable probabilistic audio creativity from text or audio inputs. This paper introduces initial work consider digital aspects of audible infrastructures (Devine and Boudreault-Fournier, 2019). Drawing on the concept of platforms becoming infrastructure (Plantin et al, 2016), a static reading of Stability AI open audio is used to explicate the relationships within it and understand the computational culture that creates the objects that we study. Algorithmic and dataset genealogies contextualise Stability’s model and relationships within and without the ecosystem. Set within the context of the recent boom in audio models, this tracing and modelling sheds light on platforms becoming infrastructural devices through models, datasets, and algorithms. The relations point to sound as a signal through algorithmic re-use and adaptation. A distant reading of the Maestro and Free Music Archive dataset metadata explores the terms and composers as tokens using counts, networks, and text processing. The terms’ simplicity and limitations affect what can be used as an audio input token stream to interact with the sound signal. The readings are used to explore the limitations in both the model and the generated audio. As platforms become infrastructural ecologies that enable creative and knowledge work, limitations and control are embedded as text is restructured into audio. The introduced methods develop a medium-specific digital humanities approach to studying generative AI platforms in computational music culture.