Voices in the auditory modality, like faces in the visual modality, are the keys to person recognition. This fMRI experiment investigated the neural organisation of voice categories using a voice-training paradigm. Voice-morph continua were created between two female Hungarian speakers' voices saying six monosyllabic Hungarian words, one continuum per word. Listeners were trained to categorize the middle part of the continua as one voice. This trained voice category was associated with a face. Twenty-five listeners were tested twice with a one-week delay. To induce shifts in the trained category, listeners received feedback on their judgments such that the trained category was associated with different voice-morph intervals each week, allowing within-subject manipulation of whether stimuli corresponded to a trained voice-category centre, to a category boundary or to another voice. FMRI tests each week were preceded by eighty minutes training distributed over two consecutive days. The tests included implicit and explicit categorization tasks. Voice and face selective areas were defined in separate localizer runs. Group-averaged local maxima from these runs were used for small-volume correction analyses. During implicit categorization, stimuli corresponding to trained voice-category centres elicited lower activity than other stimuli in voice-selective regions of the right STS. During explicit categorization, stimuli corresponding to trained voice-category boundaries elicited higher activity than other stimuli in voice-selective regions of the right VLPFC. Furthermore, the unimodal presentation of voices that are more associated with a face may elicit higher activity in visual areas. These results map out the way voice categories are neurally represented.