We recently developed a method to estimate speech-driven spectrotemporal receptive fields (STRFs) using fMRI. The method uses spectrotemporal modulation filtering, a form of acoustic distortion that renders speech sometimes intelligible and sometimes unintelligible. Using this method, we found significant STRF tuning only in classic auditory regions throughout the superior temporal lobes. However, our analysis was not optimized to detect small clusters of tuned STRFs as might be expected in non-auditory regions. Here, we re-analyze our data using a more sensitive multivariate procedure, and we identify STRF tuning in non-auditory regions including the left dorsal premotor cortex (left dPM), left inferior frontal gyrus (LIFG), and bilateral calcarine sulcus (calcS). All three regions responded more to intelligible than unintelligible speech, but left dPM and calcS responded significantly to vocal pitch and demonstrated strong functional connectivity with early auditory regions. However, only left dPM’s STRF predicted activation on trials rated as unintelligible by listeners, a hallmark auditory profile. LIFG, on the other hand, responded almost exclusively to intelligible speech and was functionally connected with classic speech-language regions in the superior temporal sulcus and middle temporal gyrus. LIFG’s STRF was also (weakly) able to predict activation on unintelligible trials, suggesting the presence of a partial ‘acoustic trace’ in the region. We conclude that left dPM is part of the human dorsal laryngeal motor cortex, a region previously shown to be capable of operating in an ‘auditory mode’ to encode vocal pitch. Further, given previous observations that LIFG is involved in syntactic working memory and/or processing of linear order, we conclude that LIFG is part of a higher-order speech circuit that exerts a top-down influence on processing of speech acoustics. Finally, because calcS is modulated by emotion, we speculate that changes in the quality of vocal pitch may have contributed to its response.