This study investigated how synthetic voice characteristics influence human learning in AI-mentored environments. This study uses acoustic theory and dual-stream cognitive models to examine two key voice dimensions: voice production (conventional vs. personalized synthetic voices) and voice transmission (radiation vs. conduction). Through a controlled laboratory experiment involving a maze-based experiential learning task in Minecraft, the study finds that personalized synthetic voices enhance learning when transmitted through radiation but hinder learning when delivered via conduction. These effects are explained by processing consistency, that is, how well the mode of delivery aligns with the users’ cognitive processing style. Radiation fosters deliberative thinking, benefiting from external-sounding voices, whereas conduction evokes self-talk, which may lead to misattribution and fast, shallow processing. This research extends information systems and human-computer interaction literature by showing how acoustic voice features shape cognition, engagement, and performance in AI-assisted learning.