Listening Ahead: Beta-Mediated Prediction in the Brain Through Four Modelling Lenses

When we listen to speech, the brain constantly anticipates what comes next. Recent work suggests that beta-band oscillations (~13 to 30 Hz) carry these predictions, but their effect on behaviour may stay hidden when listening is easy. The question this project tackles is what beta activity actually looks like in the brain during clear, natural speech, and what mechanisms organise it. The project builds on two recent papers from our group: a theoretical framework that classifies language-brain modelling approaches into four families (Bouton et al., 2026, Imaging Neuroscience), and a computational model that predicts specific neural signatures of beta-mediated prediction (Platonova, Dogonasheva, Giraud, Bouton, 2026). Both call for a direct empirical test on naturalistic neural data, which is what this internship will do. The student will use the open Podcast ECoG dataset (Zada et al., 2025, ds005574 on OpenNeuro: 30 minutes of natural narrative listening across 9 patients, with pre-aligned word and phoneme transcripts, syntactic annotations, and GPT-2-XL embeddings) to ask: what mechanisms organise beta activity in the cortical language network during natural narrative listening? Each of the four model families of the taxonomy will be applied to the same recordings: - multivariable static: temporal response functions per electrode - multivariate static: banded ridge regression with language-model embeddings and RSA - dynamical multivariable: spectral Granger and theta-beta phase-amplitude coupling - dynamical multivariate: HMM-MAR on multichannel beta Profile. M2 in cognitive neuroscience, computational neuroscience, or a related quantitative discipline. Strong interest in language and oscillatory dynamics. Required: Python (MNE-Python an asset), comfort with signal processing and statistical modelling. No prior ECoG experience required: preprocessed derivatives are provided with the dataset. What the student will gain. Hands-on fluency with every major class of language-brain modelling approach (encoding models, RSA, directed connectivity, latent dynamics) on real intracranial data, on a question the field is actively asking. The project is designed with a clear path to a co-authored publication and a strong methodological foundation for a PhD. To apply. Send a CV and a short motivation paragraph to sophie.bouton@pasteur.fr.