Adults can learn to identify nonnative speech sounds with training, albeit with substantial variability in learning behavior. Increases in behavioral accuracy are associated with increased separability for sound representations in cortical speech areas. However, it remains unclear whether individual auditory neural populations all show the same types of changes with learning, or whether there are heterogeneous encoding patterns. Here, we used high-resolution direct neural recordings to examine local population response patterns, while native English listeners learned to recognize unfamiliar vocal pitch patterns in Mandarin Chinese tones. We found a distributed set of neural populations in bilateral superior temporal gyrus and ventrolateral frontal cortex, where the encoding of Mandarin tones changed throughout training as a function of trial-by-trial accuracy ("learning effect"), including both increases and decreases in the separability of tones. These populations were distinct from populations that showed changes as a function of exposure to the stimuli regardless of trial-by-trial accuracy. These learning effects were driven in part by more variable neural responses to repeated presentations of acoustically identical stimuli. Finally, learning effects could be predicted from speech-evoked activity even before training, suggesting that intrinsic properties of these populations make them amenable to behavior-related changes. Together, these results demonstrate that nonnative speech sound learning involves a wide array of changes in neural representations across a distributed set of brain regions.