Have you heard it? A computational-behavioral study of auditory false memories

False memory studies are primarily based on the Deese-Roediger-Dermott (DRM; Roediger & McDermott, 1995) paradigm, which demonstrates memory distortions using lists of semantically related words. While this traditional method can explain memory distortions semantically, it falls short in explaining errors arising from the physical and acoustic properties of sounds. This study extends the false memory framework to the auditory domain with the use of a deep learning model known as YAMNET. Thirteen sound categories consisting of natural sounds (e.g., rain, dog) were selected for the study, excluding human voice to avoid any linguistic or speech processing that could interfere with the results. Recognition lists were structured with target sounds, similar (close) sounds, and different (distant) sounds for each sound category selected. During the study, participants completed three phases: the encoding phase, the distraction task, and the recognition test. The False Memory Generator, which utilizes vector space models, was used to automate the process of developing structured lists and to create DRM lists that are not based on verbal cues. It was expected that a higher rate of misrecognition would be found for similar compared to dissimilar sounds, thus supporting the hypothesis that acoustic/perceptual similarity modulates false memory formation. This study aims to present a data-driven approach to investigating false memories that goes beyond traditional language-based approaches, as well as the versatility of the FMG in auditory memory research.