This paper describes the techniques and approaches developed at IIIT Hyderabad for building synthetic voices in Blizzard 2008 speech synthesis challenge. We have submitted three different voices: English full voice, English ARCTIC voice and Mandarin voice. Our system is identified as D. In building the three voices, our approach has been to experiment and exploit syllable-like large units for concatenative synthesis. Inspite of large database supplied in Blizzard 2008, we find that a back-off strategy is essential in using syllable-like units. In this paper, we propose a novel technique of approximate matching of the syllables as back-off technique for building voices.