The human superior temporal sulcus (STS) is responsive to visual and auditory information, including sounds and facial cues during speech recognition. We investigated the functional organization of STS with respect to modality-specific and multimodal speech representations. Twenty younger adult participants were instructed to perform an oddball detection task and were presented with auditory, visual, and audiovisual speech stimuli, as well as auditory and visual nonspeech control stimuli in a block fMRI design. Consistent with a hypothesized anterior-posterior processing gradient in STS, auditory, visual and audiovisual stimuli produced the largest BOLD effects in anterior, posterior and middle STS (mSTS), respectively, based on whole-brain, linear mixed effects and principal component analyses. Notably, the mSTS exhibited preferential responses to multisensory stimulation, as well as speech compared to nonspeech. Within the mid-posterior and mSTS regions, response preferences changed gradually from visual, to multisensory, to auditory moving posterior to anterior. Post hoc analysis of visual regions in the posterior STS revealed that a single subregion bordering the mSTS was insensitive to differences in low-level motion kinematics yet distinguished between visual speech and nonspeech based on multi-voxel activation patterns. These results suggest that auditory and visual speech representations are elaborated gradually within anterior and posterior processing streams, respectively, and may be integrated within the mSTS, which is sensitive to more abstract speech information within and across presentation modalities. The spatial organization of STS is consistent with processing streams that are hypothesized to synthesize perceptual speech representations from sensory signals that provide convergent information from visual and auditory modalities.