Dynamic measures of performance often reveal graded features of language comprehension, even when task demands are categorical in nature. Here, we examine online measures of arm-movement trajectories in a word recognition task. Stimuli comprised two vowel continua of natural-sounding synthesized words: A duration continuum ("pen" to "pan") and a spectral continuum ("pin" to "pen"). Analyses of word identification based on the final decision reached on each trial reveal categorical behavior in vowel perception along each continuum. However, measures of mouse movements toward the unselected alternative suggest that stimuli at the "long" end of the duration continuum are temporarily perceived to be more consistent with the shorter vowel (/E/) even for stimuli eventually classified as the longer vowel (/ae/) on 100% of trials. This suggests that the process of speech sound categorization favors rapid action based on the stimulus most consistent with the currently available input, even when duration itself is the distinguishing cue in reaching a final decision.