Latent Field

Gabriele de Seta, 2025

A playback interface for exploring the latent field of ChatGPT's text-to-speech model. All samples are recordings of ChatGPT responses in which the text-to-speech model is nudged towards outputting non-verbal sounds, resulting in the appearance of glitches that seem to point towards outliying areas of the training data represented in the model's latent space. It is likely that those include the remainder of unwanted noises, interjections, pauses and ambient sounds ingested during training alongside samples of natural language. A double act of recording is at play here: ChatGPT's text-to-speech output is recorded as the synthesis of abstracted patterns derived from recordings of human speech acts. Playing them back in a new randomized order allows the listener to explore this latent field on their own terms.