Multivariate pattern analysis of input and output representations of speech
Markiewicz, Christopher Johnson
MetadataShow full item record
Repeating a word or nonword requires a speaker to map auditory representations of incoming sounds onto learned speech items, maintain those items in short-term memory, interface that representation with the motor output system, and articulate the target sounds. This dissertation seeks to clarify the nature and neuroanatomical localization of speech sound representations in perception and production through multivariate analysis of neuroimaging data. The major portion of this dissertation describes two experiments using functional magnetic resonance imaging (fMRI) to measure responses to the perception and overt production of syllables and multivariate pattern analysis to localize brain areas containing associated phonological/phonetic information. The first experiment used a delayed repetition task to permit response estimation for auditory syllable presentation (input) and overt production (output) in individual trials. In input responses, clusters sensitive to vowel identity were found in left inferior frontal sulcus (IFs), while clusters responsive to syllable identity were found in left ventral premotor cortex and left mid superior temporal sulcus (STs). Output-linked responses revealed clusters of vowel information bilaterally in mid/posterior STs. The second experiment was designed to dissociate the phonological content of the auditory stimulus and vocal target. Subjects were visually presented with two (non)word syllables simultaneously, then aurally presented with one of the syllables. A visual cue informed subjects either to repeat the heard syllable (repeat trials) or produce the unheard, visually presented syllable (change trials). Results suggest both IFs and STs represent heard syllables; on change trials, representations in frontal areas, but not STs, are updated to reflect the vocal target. Vowel identity covaries with formant frequencies, inviting the question of whether lower-level, auditory representations can support vowel classification in fMRI. The final portion of this work describes a simulation study, in which artificial fMRI datasets were constructed to mimic the overall design of Experiment 1 with voxels assumed to contain either discrete (categorical) or analog (frequency-based) vowel representations. The accuracy of classification models was characterized by type of representation and the density and strength of responsive voxels. It was shown that classification is more sensitive to sparse, discrete representations than dense analog representations.