Advanced MEG Analysis of Auditory and Linguistic Encoding in Spoken Language Processing

In this work, we explore brain responses related to language processing using neural activity elicited from auditory stimuli and measured through Magnetoencephalography (MEG). We develop audio (i.e. stimulus)-MEG encoders using both time-frequency decompositions and latent representations based on wav2vec2 embeddings, and text-MEG encoders based on CLIP and GPT-2 embeddings, to predict brain responses from audio stimuli only. The analysis of MEG signals reveals a clear encoding pattern of the audio stimulus within the MEG data, highlighted by a strong correspondence between real and predicted brain activity. Brain regions where this correspondence was highest were lateral (vocal features) and frontal (textual features from CLIP and GPT-2 embeddings).