Il ruolo della somiglianza uditiva nel fenomeno dei falsi ricordi.

Semantic memory has traditionally been investigated through paradigms that, although very reliable, emphasize its associative and linguistic structure. The Deese-Roediger-Mcdermott (DRM) task has represented one of the most reliable and influential experimental tools for investigating false memories. Extensive research has been conducted on the role of surface-level and phonological similarity in the DRM, showing that they could elicit false recalls and recognitions, considerably less attention has been dedicated towards the study of perceptual similarity per se, independently from linguistic content. Notably, the role of auditory similarity per se, which could be defined as estimate of how close two sounds are, based on many variables ranging from low-level (e.g. rhythm and pitch) to high-level (e.g. whether the sound was produced by a guitar or it’s speech), has remained largely unexplored. The present thesis investigates whether auditory similarity alone is able to produce false memories in the DRM paradigm, using non-linguistic stimuli. To achieve this, it was implemented a completely data-driven approach. Sound stimuli are selected from AudioSet and turned into mel- spectrograms; these are in turn, represented within a vectorial embeddings space by a convolutional neural network (CNN): YAMNet, which extracts auditory feature from them. DRM lists are automatically created using False Memory Generator (FMG), ensuring that these lists’ composition is determined through auditory similarity rather than human association norms. Crucially, all stimuli excluded human speech, allowing for the disentanglement between auditory and linguistic similarity. Forty participants completed the DRM task; in the recognition phase they were presented items varying in auditory similarity from the Studied Items. The data is analysed using generalized linear mixed models (GLMMs) to assess whether the auditory similarity would predict false recognition rates. Results show that Critical Lures were falsely recognized much more frequently than Unrelated Lures, and that false recognition rates increased as a function of auditory similarity between Studied Items and Lures. These findings provide the first evidence that auditory similarity per se, isolated from linguistic aspects, can drive false memory formation. These results extend the already growing evidence that perceptual similarity contributes to the phenomenon of false memory and in turn, to semantic memory. Key words: Semantic memory, DRM paradigm, Auditory similarity, CNNs, FMG.