Abstract
The remarkable speed and immediacy of human speech comprehension depends on the dynamic integration of acoustic-phonetic cues from the speech input with multi-dimensional contextual constraints. To understand the neurocomputational infrastructure that supports these processes, we tested where the critical brain regions are, when these different inputs are integrated, how different cues and constraints are neurally represented and what is the content of connectivity processes across the network. To do these, we combine EMEG measures of real-time brain activity with advanced NLP models, imaging methods (GCA, RSA) and network approaches (ICA and lasso). We uniquely identify a LH fronto-temporal network that integrates acoustic-phonetic cues and contextual constraints to support the identification of lexical form and meaning in the first 200. This integration process is not visible at word-onset, which suggested that bottom-up constraints are necessary to set the representational geometry of an analysis space with which semantic constraints can interact.
