In this paper, we will discuss the selection of synthesis unit for Mandarin speech synthesis. To choose the synthesis unit which could be easily segmented and concatenated, and produce natural and intelligent speech, we proposed the synthesis unit of Coarticulatory Demi-syllable segmented from di-syllable words. We found that spectral continuity of concatenating demisyllable still kept good. By subjectively listening, testers felt that the Coarticulatory Demi-syllable produced more natural and intelligent synthetic speech than monosyllable, but less than synthesis-by-analysis speech.