Green MIR? Investigating Computational Cost of Recent Music-Ai Research in ISMIR

The environmental footprint of Generative AI and other Deep Learning (DL) technologies is increasing. To understand the scale of the problem and to identify solutions for avoiding excessive energy use in DL research at communities such as ISMIR, more knowledge is needed of the current energy cost of the undertaken research. In this paper, we provide a scoping inquiry of how the ISMIR research concerning automatic music generation (AMG) and computing-heavy music analysis currently discloses information related to environmental impact. We present a study based on two corpora that document 1) ISMIR papers published in the years 2017–2023 that introduce an AMG model, and 2) ISMIR papers from the years 2022–2023 that propose music analysis models and include heavy computations with GPUs. Our study demonstrates a lack of transparency in model training documentation. It provides the first estimates of energy consumption related to model training at ISMIR, as a baseline for making more systematic estimates about the energy footprint of the ISMIR conference in relation to other machine learning events. Furthermore, we map the geographical distribution of generative model contributions and discuss the corporate role in the funding and model choices in this body of work.