The Tensed Computer Improviser and the workmanship of risk: Small-dataset machine learning and real-time interactive musical improvisation

Abstract Generative artificial intelligence (AI) has entered music overwhelmingly by way of large industrial models trained on vast corpora, in which the outcome is fixed in advance by an apparatus that the musician neither controls nor understands. This article argues that the essential machine-learning operations of dataset assembly, feature extraction and model fitting can instead be utilised transparently as part of an artist’s creative practice. Looking to David Pye’s concepts of workmanship of certainty and workmanship of risk , this article examines the design of the Tensed Computer Improviser (TCI), an open-source machine improviser built in Max with the Fluid Corpus Manipulation (FluCoMa) toolkit, which trains models from a small audio corpus of the user’s own choosing, with no pretrained component and no external training stage. Situating the TCI within a genealogy of machine improvisation using small data, the article analyses its design through the vocabulary of craft where the corpus curation and training are integrated into the creative act and the characteristic errors of small data are welcomed as tool-marks.