The present thesis is dedicated to studying the human language connectome by combining diffusion-weighted magnetic resonance imaging (dMRI) and high-dimensional predictive algorithms to isolate its relevant connections amongst the broader white matter feature space. According to the classical “Broca-Wernicke” and contemporary “Dual-Stream” neurobiological models of language, the fluent production and repetition of speech relies on a distinct and lateralized neuroanatomical circuit. Where these models diverge is in their neuroanatomical explanations, such that the former attributes these functions to the posterior inferior frontal gyrus (i.e., Broca’s area) and its disconnection from the arcuate fasciculus while the latter emphasizes a cortico-centric “dorsal stream” network where repetition processing occurs by way of a sensorimotor node in sylvian parieto-temporal cortex (called “area Spt”). Experiment 1 of this thesis tested the assumption that the speech production system is scaffolded by a lateralized set of connections by performing whole-brain inter-hemispheric white matter comparisons between left- and right-hemisphere dominant subjects as determined by a clinically-indicated Intracarotid Amytal (Wada) Test. Indeed, this semi-automated analysis reconstructed a lateralized network for speech production that was underscored by its convergence of pathways to Broca’s area and whose white matter properties successfully enabled Wada-concordant classifications using a ridge-logistic regression model. Experiment 2 used dMRI collected by the Human Connectome Project to perform multi-shell fiber tractography of Broca’s area in order to segment the underlying sub-circuits connecting to this classical region. Experiment 3 linked connectivity directly with behavior using a connectome-based lesion-symptom mapping approach in a large dMRI sample acquired from stroke patients. Over two-thousand distinct connections were mapped and a LASSO-regularized latent projection-based regression algorithm was implemented to automatically isolate the subset of connections predictive of speech repetition performance. We found that the embedded feature selection algorithm identified a local set of sylvian parieto-temporal connections that could make accurate out-of-sample predictions of repetition performance - a result which fuses the claims made by the “Broca-Wernicke” and “Dual-Stream” models regarding the anatomic delineation of the repetition circuit. Together, these studies demonstrate the utility in using dMRI and machine-learning algorithms to understand the anatomy of eloquent functional networks while accommodating the collinearities that are characteristic of brain data.

A densely interconnected patchwork of cortical areas tiles the brain, balancing the integration and segregation of functions. The eminent strategy for charting the functions of this landscape has involved pinning tightly controlled behavioral tasks onto large and often isolated patches of cortex in piecemeal. Recent developments have made it possible to study the organization of this rich functional landscape from the bottom-up, by mapping the connectivity networks that underpin brain areas situated within larger functional regions. At the same time, the availability of large databases that aggregate published functional neuroimaging work make it possible to infer the functions of such connectivity-defined areas from a much broader behavioral space than can be accessed by any individual study. In the first study of the present work, this bottom-up approach was leveraged to elucidate functional organization around the planum temporale (PT), an anatomical region that has been shown to be critical to speech production, but which has also been associated with a variety of other functions in functional neuroimaging studies. Functional connectivity estimated from multiecho resting state data (N =137) was used to map areas within the posterior perisylvian zone that functional neuroimaging studies have associated with the PT. A deliberate parcellation of this zone was produced by forming cluster ensembles from complementary unsupervised learning algorithms that produced reliable parcellations (12 clustering approaches were tested for their reliability). The resulting parcellation was well-aligned with observer-independent architectonic work, and even outperformed several popular atlases when used to classify external data. Registering the networks and areas within this parcellation to published functional neuroimaging work revealed that the functional zone associated with the PT was comprised of multiple speech areas, including an area involved in rehearsal and an area involved in production. Further, multiple areas were associated with receptive speech, auditory, and sensorimotor processing. At a broad level, a second study tested the extent to which there can be agreement between areas defined from the bottom-up based on connectivity data, and functional areas embedded in published neuroimaging work. That is, a parcellation of temporal and inferior parietal cortex based on structural connectivity (N = 70) was compared to a parcellation resulting from the same clustering approach adapted to brain-behavior associations (i.e., meta-analyses of behaviors). This functional parcellation carved out areas with similar behavioral profiles and was implemented in the service of localizing areas involved in speech perception. Prior meta-analytic studies with the same goal have used sample sizes now known to be insufficient, produced incompatible results, and failed to consider whether speech responsive areas are involved in other functions. Clustering was performed over lower-dimensional functional networks extracted from brain-behavior associations. Auditory and speech processes were robustly distinguished irrespective of how many functional networks were defined. A more optimal set of functional networks showed evidence of a posteriorly directed speech processing hierarchy that implicated an area in ventral superior temporal gyrus, posterior to auditory cortex, in speech sound analysis. The functional parcellation isolated areas along this speech hierarchy and captured the extent to which they were shared across functional networks. Overall, there was significant spatial correspondence between the functional and structural parcellations, and the parcel associated with speech sound processing showed the best fit to structural connectivity data. Structural data indicated that this parcel could be distinguished based on its strong connectivity along the middle longitudinal fasciculus. A third study investigated a more targeted hypothesis about the role of pMTG in error correction. The results of this study helped better characterize the function of one of two areas from the functional parcellation presented in the second study that were both associated with the pMTG, but difficult to functionally distinguish in a precise way. Previous work has demonstrated suggestive evidence of internal error detection and correction in left posterior middle temporal gyrus (pMTG) on the basis that this area tended toward showing a stronger response when potential speech errors (i.e., not realized) are biased towards nonwords compared to words. Building on this prior work, an experiment was designed introducing novel tongue twister stimuli that attempted to further tax internal error correction and detection mechanisms by biasing potential speech errors towards taboo words. Although the pMTG demonstrated significantly greater response to taboo words than neutral words in a large sample of participants (N=40), other areas also showed an effect. Behavioral associations in the wider neuroimaging literature were used to better understand the functional significance of activity outside the pMTG, demonstrating that only activity within this area was likely to signal word-level processing, and therefore explain the internal error correction mechanism investigated. Taken together, these studies present a more nuanced description of the functional and structural networks underpinning the speech system and point to more specific functions that these networks may support.

We investigated a new computational model of speech production, the Semantic-Lexical-Auditory-Motor Model (SLAM), which was designed to test a critical assumption of the Hierarchical State Feedback Control theory (Hickok, 2012): specifically, that speech production relies on the coordination of dual representations in auditory and motor cortices, with auditory representations serving as targets during speech planning. The computational details are based on the interactive, two-step lexical retrieval model of Foygel and Dell (2000); our novel architecture allows us to predict the consequences of different patterns of damage among the proposed speech representations. The additional model structure is expected to better explain conduction aphasia in particular. We analyzed archived picture naming data from 255 people with aphasia in Philadelphia, PA, in addition to new data from another 95 people with aphasia in Columbia, SC. We found that the SLAM model made adequate predictions generally, and it did improve the fit to data from conduction patients specifically and in the expected manner. We also analyzed neuroanatomical data in the form of lesion masks standardized to a template for 83 of the participants from the SC cohort. Although we were unable to replicate a study to localize brain regions where damage leads to a significant increase in particular error types, a behavioral comparison of the different cohorts revealed the potential sampling variability that exists in the aphasia population. Next, we developed a Bayesian approach to estimating the parameters of the lexical network, providing a more comprehensive assessment of the model’s quality. Additionally, we simulated word and nonword repetition tasks with our network, generating new predictions for a subset of 28 people with aphasia and unimpaired hearing. We found that the 4-parameter SLAM model could simultaneously find good fits for the frequencies of 6 naming and 3 nonword repetition response types while also correctly predicting a novel set of 6 word repetition response types. The conduction patients' data was best explained by strong lexical-auditory and weak auditory-motor connections. Our results demonstrate that the assumption of coordinated speech representations in auditory and motor cortices can lead to viable predictions of speech production behavior in aphasia.

Human speech processing (perception and in some cases production) is approached from three levels. At the top level, I investigate the role of the motor system in top-down processing and decision-making during speech perception. At the middle level, I investigate the mechanisms underlying integration of auditory and visual speech for both perception and production of speech. At the bottom level, I investigate the organized representation of temporal modulations in sound, with an eye toward structure that may provide insight into how speech sound representations are built. The primary investigative techniques throughout are auditory and visual psychophysics and functional MRI (sometimes combined). The main findings of the investigations can be summarized briefly as follows. First, the motor system does not participate meaningfully in speech perception. Rather, speech motor activity is modulated by taxing decision-level mechanisms in laboratory speech tasks. Second, discrete visual features appear to be extracted from visual speech signals and integrated with auditory speech representations in the superior temporal sulcus (STS). Results are equivocal with respect to the level of processing at which this occurs, although speculation is provided. Also, there are dedicated sensorimotor integration networks for visual speech. Third, slow temporal modulations in sound are represented in an auditory-cortical place code that magnifies the expression of modulations within the range that is most common in natural speech (4-16 Hz).

This dissertation critically examines extant hypotheses on the syntax-brain relationship, particularly proposals concerning Broca's area, and presents empirical data in efforts to localize syntactic operations in the brain. Chapter 2 of the dissertation reviews arguments for and against a role for Broca's area in syntax, and presents an fMRI experiment using a construction called backward anaphora in which the activity in the pars triangularis of Broca's area is shown to pattern with how sentences are processed (i.e., whether the construction involves active processing), and not their syntactic properties (i.e., activity is not contingent on Movement constructions). Chapter 3 extends Dehaene & Cohen's neuronal recycling hypothesis (Dehaene & Cohen, 2007) to language function in Broca's area to account for sentence-selective activations in this region of cortex (Fedorenko et al., 2012). In particular, the discussion in Chapter 3 focuses on language- or sentence-specific working memory (Caplan & Waters, 1999), and how language-specific working memory may develop out of domain-general working memory. Chapter 4 presents an fMRI experiment aimed at uncovering the neural basis of syntax using a "syntactic perturbation" technique during overt sentence production. In addition, the experiment specifically examined the response profile of Broca's area and another region classically implicated in structural processing, the anterior temporal lobe (ATL), to our novel manipulation targeting syntactic structure. The results, when compared to perturbation in a control condition of unstructured word lists, revealed preferential activation for syntactic perturbation in networks previously implicated in motor control (Tourville et al., 2008) and action inhibition (Aron et al., 2014), including the basal ganglia, thalamus, and right inferior frontal gyrus, suggesting that (i) sentence production allows greater phonological/articulatory planning than list production, and (ii) syntax may rely on networks similar to basic motor control. However, effects at the syntax level are conflated with effects at lower motor levels in our experiment. Broca's area and the ATL did not exhibit effects consistent with syntactic processing, although our results extended the basic finding of increased activation to sentences compared to word lists previously found in the ATL (Mazoyer et al., 1993). Finally, Chapter 5 concludes the dissertation.

There are divergent claims concerning the broad cortical organization of speechrecognition. One model holds that speech perception and comprehension is governed by a left lateralized anterior temporal lobe (ATL) pathway. Another model argues that bilateral superior temporal regions are critically important, and, in fact, represent a lower level of processing that drives ATL activation in a bottom up fashion. These models were tested in a series of auditory fMRI experiments that gradually investigated lower levels of speech analysis. The experiments contrasted listening to clear monosyllabic words, pseudowords, sentences, and word lists with unintelligible spectrally rotated and time-reversed speech. In the first experiment, posterior temporal regions did not respond differentially to sentence versus word list stimuli, consistent with the idea that bilateral regions of the superior temporal plane support speech recognition at a lower (perhaps phonological) level. An area of the ATL centered around the superior temporal sulcus (STS) was activated more for sentences than word lists, indicating that the region may be involved in sentence-level operations. In the second experiment, this same region in the left hemisphere was activated more by monosyllabic words than rotated words. This suggests that the anterior focus is not exclusively attributable to sentence-level operations. In the third experiment, lexical status was found to differentially modulate anterior and posterior STS regions. There was more activation in the aSTS bilaterally for words than pseudowords, but these conditions did not lead to activation differences in the posterior region. It appears that anterior temporal speech-selective regions respond to lexical-semantic aspects of speech, whereas posterior temporal speech-selective areas are coding lower level phonemic information.