Neural activity flows through cortical subnetworks during speech production.

Speech production entails several processing steps that encode linguistic and articulatory structure, but whether these computations correspond to spatiotemporally discrete patterns of neural activity is unclear. To address this issue, we used electrocorticography to directly measure the brains of neurosurgical participants performing an interactive speech paradigm. We observed a broad range of cortical modulation profiles, and subsequent clustering analyses established that responses comprised distinct classes associated with sensory perception, planning, motor execution, and task-related suppression. These activity classes were also localized to separate neural substrates, indicating their status as specialized networks. We then parsed dynamics in the planning and motor networks using unsupervised dimensionality reduction, which revealed subnetworks that were sequentially active throughout preparation and articulation. These results therefore support and extend a localizationist model of speech production where cortical activity 'flows' within and across discrete pathways during language use.