Una prospettiva critica sui dataset per la sintesi text-to-audio

Text-to-audio synthesis (TTA) promises sound generation mediated solely by natural language. Like all generative deep learning techniques, it derives its world memory and semantic boundaries from the dataset with which it is trained. Following a humanistic tradition of critical algorithm analysis, this article examines the datasets employed by various text-to-audio synthesis models and how they are designed and manipulated by researchers. Datasets are not inert informational structures, but socio-technical objects involving specific symbolisations mediated by cultural, po litical, and industrial perspectives. A comparative study of label datasets, caption datasets, and algorithmically augmented datasets reveals the technical and ethical limitations associated with a quantitative approach to data collection. In contrast to a purely computational dataset evaluation paradigm, a qualitative analysis methodology is proposed to interpret the connotative capability of models and suggest alternative practices in the collection of information.