AI SONG CONTEST: HUMAN-AI CO-CREATION IN SONGWRITING Cheng-Zhi Anna Huang1 Hendrik Vincent Koops2 Ed Newton-Rex3 Monica Dinculescu1 Carrie J. Cai1 1 Google Brain 2 RTL Netherlands 3 ByteDance annahuang,noms,
[email protected], h.v.koops,
[email protected] ABSTRACT experiences were enabled by major advances in machine learning and deep generative models [18, 66, 68], many of Machine learning is challenging the way we make music. which can now generate coherent pieces of long-form mu- Although research in deep generative models has dramati- sic [17, 30, 39, 55]. cally improved the capability and fluency of music models, Although substantial research has focused on improving recent work has shown that it can be challenging for hu- the algorithmic performance of these models, much less is mans to partner with this new class of algorithms. In this known about what musicians actually need when songwrit- paper, we present findings on what 13 musician/developer ing with these sophisticated models. Even when compos- teams, a total of 61 users, needed when co-creating a song ing short, two-bar counterpoints, it can be challenging for with AI, the challenges they faced, and how they lever- novice musicians to partner with a deep generative model: aged and repurposed existing characteristics of AI to over- users desire greater agency, control, and sense of author- come some of these challenges. Many teams adopted mod- ship vis-a-vis the AI during co-creation [45]. ular approaches, such as independently running multiple Recently, the dramatic diversification and proliferation smaller models that align with the musical building blocks of these models have opened up the possibility of leverag- of a song, before re-combining their results.