Unit Selection based speech synthesis systems (USS) require accurate labeling of units. Accurate segmentation of speech waveforms manually is a laborious task. Syllable-based systems for Indian languages use a group delay based approach for semi-automatic segmentation of speech waveforms into syllables. This performance of the group delay based algorithm is poor when the syllables contain fricatives, nasals and unvoiced stop consonants. This paper proposes a modification to the algorithm that exploits the properties of these types of units to reduce errors. In particular, the ratio of energy in the high frequency bands to low frequency bands is used as a cue to segment the speech signal.