Oral communication is the most important way for delivering information in our daily life. Unfortu-nately, the quality of such communication can be degraded by 1) speech disorders (e.g. dysarthria) and 2) surrounding environments (e.g. noise or reverberation). Style conversion is a technology that modiﬁes the source speaking style of a speaker to sound like a more intelligible target speak-ing style of either the same or diﬀerent speaker. In the dissertation, I consider new machine learning based-approaches for style conversion. Inspired by the intelligibility gain of clear (CLR) speaking style over habitual (HAB) speaking style, I propose several HAB-to-CLR spectral mappings approaches for intelligibility improvement.