Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Cortical encoding of probabilistic temporal predictions during speech processing
Poster Session A, Wednesday, September 30, 11:00 am - 1:00 pm, Wangari Maathai
Laure Deyna1, Philippe Albouy, Agnes Trebuchon, Daniele Schön, Benjamin Morillon, Pierre Hieu Guilleminot; 1Aix-Marseille University, Inserm, Institut de Neurosciences des Systèmes (INS), Marseille, France, 2CERVO Brain research center, Laval University, Québec, Canada, 3Centre Hospitalier Universitaire de Québec, Laval University, Québec, Canada, 4APHM, Timone Hospital, Epileptology and Cerebral Rhythmology Department, Marseille, France
While speech unfolds across multiple temporal scales — from phonemes to syllables and words — it has traditionally been temporally characterized by their mean occurrence rates. We show that at each linguistic level speech exhibits complex temporal patterns that are both predictable and that this probabilistic predictability is represented in neural activity. Using large French and English speech corpora, we trained models of increasing complexity to predict linguistic onsets. Recurrent neural networks (RNNs) outperformed mean-rate and hazard-rate models, and RNNs jointly modelling phoneme, syllable and word occurrences performed best, revealing that temporal dependencies across hierarchical linguistic levels shape the temporal structure of speech. The dominant level of predictability - syllables in French and words in English - reflected the distinct rhythmic typology of syllable-timed and stress-timed languages. Intracranial EEG further showed that auditory and motor cortices encode these probabilistic temporal predictions during natural speech processing, demonstrating that the brain dynamically tracks the hierarchical temporal structure of speech.
Topic Areas: Speech Perception, Multisensory or Sensorimotor Integration