We propose a segment-concatenation algorithm which reduce perceived distortion caused by segment concatenation for a segment-based speech synthesizer. This algorithm concatenates six types of phonetic segments along the transient part of speech rather than the steady part, where humans have a keen sense of spectral distortion. This concatenation method enables a segment-based synthesizer to produce a smooth sound with comparatively small required storage space for the segments. We apply the algorithm to a rule-based, cepstrum-based speech synthesizer for English words. We evaluate the intelligibility of the synthetic speech through the Modified Rhyme Test (MRT)[1]. The result proved that the speech has a high intelligibility ratio of 90 percent.