Figure 1.

(A) Analyses considered the pattern of activation across several regions associated with language processing: bilateral frontal regions (inferior and middle frontal gyri), insular cortex, temporal regions (superior, middle, and transverse temporal gyri), and parietal regions (supramarginal and angular gyri). The analyses in this figure considered the classification ability of voxels in all these regions, irrespective of specific location.

(B) Behavioral performance on the phonetic categorization task, previously reported by Myers and Mesite (2014). Subjects made more /s/ responses as stimuli became more /s/-like, and their overall rate of /s/ responses was higher if they had previously encountered ambiguous stimuli in contexts where lexical information guided them to interpret the ambiguous stimulus as ‘s.’ Critically, there is still a considerable amount of trial-to-trial variability in classification of ambiguous tokens; for instance, subjects in the /s/-biased group interpreted the 40% /s/ token as /s/ approximately half the time and as /∫/ half the time.

(C) A classifier was trained to classify unambiguous stimuli as /s/ or /∫/ and was able to successfully generalize to ambiguous stimuli near the phonetic category boundary, with significantly above-chance accuracy in determining how subjects perceived an ambiguous stimulus on each particular trial (left panel). However, the classifier did not reach above-chance accuracy when ambiguous stimuli were labeled with respect to acoustics (right panel). Dark points indicate mean accuracy, and error bars indicate standard error. Light lines indicate classification accuracy by subject, with light red lines indicating subjects who had previously heard ambiguous tokens in /s/-biased contexts and light blue lines indicating those who had previously received /∫/-biased contexts.

(D) A follow-up analysis indicated that the classifier’s ability to recover trial-by-trial perceptual experiences was numerically related to individual differences in phonetic recalibration. In particular, the classifier was most accurate in classifying ambiguous tokens for subjects who showed large behavioral effects (measured as the proportion of bias-consistent responses made on ambiguous trials). Red points indicate subjects who had previously heard ambiguous tokens in /s/-biased lexical frames, while blue points indicate those who had heard ambiguous tokens in /∫/-biased contexts.