Neural representation can be induced without external stimulation, such as in mental imagery. Our previous study found that imagined speaking and imagined hearing modulated perceptual neural responses in opposite directions, suggesting motor-to-sensory transformation and memory retrieval as two separate routes that induce auditory representation (Tian & Poeppel, J Cogn Neurosci. 2013 25:1020-36). We hypothesized that the precision of representation induced from different types of speech imagery led to different modulation effects. Specifically, we predicted that the one-to-one mapping between motor and sensory domains established during speech production would evoke a more precise auditory representation in imagined speaking than retrieving the same sounds from memory in imagined hearing. To test this hypothesis, we built the function of representational precision as the modulation of connection strength in a neural network model. The model fitted the MEG imagery repetition effects, and the best-fitting parameters showed sharper tuning after imagined speaking than imagined hearing, consistent with the representational precision hypothesis. Moreover, this model predicted that different types of speech imagery would affect perception differently. In an imagery-adaptation experiment, the categorization of/ba/-/da/continuum from male and female human participants showed more positive shifts towards the preceding imagined syllable after imagined speaking than imagined hearing. These consistent simulation and behavioral results support our hypothesis that distinct mechanisms of speech imagery construct auditory representation with varying degrees of precision and differentially influence auditory perception. This study provides a mechanistic connection between neural-level activity and psychophysics that reveals the neural computation of mental imagery. Our brain processes sensory information that we receive from the environment and mediates mental activity such as imagination. How do mental and perceptual processes interact to shape our cognition? We constructed a computational model that simulated how two types of imagery - imagined speaking and imagined hearing -differentially modulated perception via two distinct neural pathways. This model further predicted a choice shift in perceptual responses to ambiguous auditory stimuli, which was confirmed in a follow-up imagery-adaptation experiment. These results suggest that parallel neural pathways for mental imagery provide distinct functions to influence perception. These findings may implicate multiple strategies for constructing the brain-computer interface.