Cortical representations of language components studied with human electrocorticography

Speech production engages a variety of subprocesses, including conceptual preparation, lexical selection, phonemic sequencing, access to the motor articulatory code, motor execution, and online auditory and somatosensory monitoring. Likewise, the perception and comprehension of speech engages similar processes in reverse order (auditory processing, phonemic extraction and sequencing, lexical access, and conceptual activation). The underlying cortical network is complex and distributed across multiple functionally-specialized anatomical regions. While these subprocesses and their associated cortical sites have largely been identified, questions remain regarding the specific representations and transformations of these language components in the human brain. In this dissertation, semantics and phoneme extraction were selected for study; electrocorticography was recorded while subjects performed one of two associated language tasks. To investigate semantics, subjects performed image naming. Regression was used to build encoding models that mapped between a semantic vector space and high gamma (HG, 70-110 Hz) power extracted from the neural signals. High gamma power served as a surrogate of neural population firing rates. These models could predict objects withheld during training with state-of-the-art accuracy. Further, cortical sites located on the basal occipitotemporal surface encoded specific semantic dimensions such as manmade vs. alive and large vs. small. To investigate phoneme extraction, subjects performed a phoneme identification task while listening to eight speech segments, artificially generated by specifying formant frequencies and transition times. These stimuli were positioned linearly: initial F2 and F3 formant frequencies were evenly spaced, with all other parameters held constant. A sharp transition boundary divided stimuli into two perceptual categories (/ba/ and /da/). These results demonstrated categorical perception: adjacent stimuli on the same side of a discrimination boundary were perceived identically while equivalently spaced stimuli straddling the boundary were not. Logistic regression decoding models using HG revealed that lateral superior temporal gyrus may mediate the transformation of spectrotemporal features to category membership. A novel technique designed to estimate neural representations from decoding accuracy patterns or model coefficients revealed dynamic transitions from early gradient to later categorical neural representations in a plurality of cortical sites. Finally, a network model was proposed to account for specific morphological aspects of these stimulus representations