Christopher Markiewicz¹, Jason Bohland²
¹Boston University, ²Boston University
Human speech is subserved by a bilateral network of cortical regions, including superior temporal, inferior parietal and inferior frontal cortex, and utilizes auditory, phonological, and motor representations of speech sounds. The precise nature of these representations is difficult to assess due to the limitations of functional neuroimaging and the complexity of speech sounds. Multi-voxel pattern analysis (MVPA) provides a tool for studying representations using the accuracy of statistical models as indicators of local information content. In this fMRI study, 14 healthy, adult English speakers were presented with systematically varying consonant-vowel-consonant syllables in a delayed repetition task. Each trial began with the auditory presentation of a syllable, and overt repetition was visually cued after a ~9s delay. Two whole-brain EPI volumes were collected per trial, timed to the peak hemodynamic responses to the stimulus onset and the production cue. Cortical surfaces were reconstructed and MVPA, using a surface-based searchlight approach, was used to generate information maps for discrete stimulus features based on the cross-validation accuracy of linear support vector machines. Maps were created for the vowel of each syllable and for syllable identity, separately based on the perception and production responses. Group-level results show left-lateralized regions in the posterior inferior frontal sulcus (IFS) and supramarginal gyrus (SMG) that significantly predicted vowel identity for perception events. For production events, bilateral superior temporal sulcus (STS) regions were identified, as well as the right temporo-parietal junction. Prediction of whole syllables highlighted anterior superior temporal regions in the perception and sensorimotor as well as superior temporal regions in production.
Keywords: MVPA, Speech