Optimal motor performance requires flexible adjustments to behavior depending on different contexts. This idea has gained wide support from behavioral and computational studies, but remains poorly understood at the neural level. Bengalese finch (Lonchura striata) song is a learned motor sequence composed of discrete elements, or syllables. Adult finches can learn to modify the pitch of individual syllables in a gradual way, supported by a well-understood cortical-basal ganglia circuit, the AFP (anterior forebrain pathway). We here tested whether adult finches can use arbitrary contextual cues to flexibly and immediately switch between different modifications of the pitch of a target syllable. We paired opposite directions of pitch reinforcement for the same syllable with different colors of cage illumination, e.g., reinforcing upward pitch shifts in orange light and downward pitch shifts in green light. After training, light switches elicited immediate adaptive changes to song that minimized aversive feedback in each context. At the behavioral level, we could dissociate two processes, an immediate cue-dependent pitch shift at the beginning of new context blocks, and a gradual reinforcement-dependent pitch adaptation within context blocks. At the neural level, we show that immediate cue-dependent pitch shifts are independent of the AFP, while lesions of this circuit abolished gradual pitch adaptation, consistent with prior studies. This indicates that rapid sensory context-dependent pitch shifts are mediated by different mechanisms than gradual pitch adjustments. Prevailing models of contextual motor adaptation argue for at least two parallel processes: one that is flexible, and sensitive to contextual information, and a second that cannot readily be associated with contextual cues and is gradually updated during motor adaptation. Our results indicate that context-dependent pitch learning in birdsong likewise involves two dissociable processes, based on different neuronal circuits. This provides a window to understand one neural implementation of dual process models in motor adaptation.