Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Persistent and evolving encoding of phonemes reflects general auditory processing

Oli Danyi Liu¹, Hao Tang², Naomi Feldman³, Sharon Goldwater²; ¹University of Edinburgh, University of Edinburgh, ²University of Edinburgh, ³University of Maryland, College Park

Presenter: Oli Danyi Liu

Speech recognition involves storing and integrating sequentially presented information. Recent work in cognitive neuroscience has identified temporal characteristics in humans' neural encoding of speech that may facilitate this process. A modeling study found similar properties in a self-supervised learning model trained on raw speech, suggesting these properties can arise without prior linguistic knowledge. In this work, we further explore the domain specificity of the same properties through testing representations of speech extracted from a model only trained on non-speech audio. The model replicated key aspects of the temporal characteristics, implying they might not be specific to speech perception, but rather features of general auditory processing.

Topic Area: Language & Communication

Extended Abstract: Full Text PDF