A quantitative method is introduced for detecting and correcting artifactual signal changes in BOLD time series data arising from the magnetic field warping caused by motion of the articulatory apparatus when speaking aloud, with extensions to detection of subvocal articulatory activity during silent reading. Whole-head images allow the large, spike-like signal changes from the moving tongue and other components of the articulatory apparatus to be detected and localized in time, providing a measure of the time of vocalization onset, the vocalization duration, and also an estimate of the magnitude and shape of the signal change resulting from motion. Data from brain voxels are then examined during the vocalization period, and statistical outliers corresponding to contamination from articulatory motion are removed and replaced by linear interpolation from adjacent, uncontaminated data points. This quantitative approach to cleansing brain time series data of articulatory-motion-induced artifact is combined with a pre-scanning training regimen that reduces gross head movement during reading aloud to the levels observed during reading silently, which can be corrected with available image registration techniques. The combination of quantitative analysis of articulatory motion artifacts and pre-scanning training makes possible a much wider range of tasks involving overt speech than are currently being used in fMRI studies of language and cognition, as well as characterization of subvocal movements of the articulatory apparatus that are relevant to theories of reading skill, verbal rehearsal in working memory, and problem solving.