Most of isolated-word speech recognition systems need to detect boundaries of utterances. In this paper, we present a new approach to accomplish this task of endpoints identification. The proposed algorithm has very low computational complexity, since the parameters used for discrimination are obtained through a re-use of calculations already made by a mel-cepstrum frontend. This endpointer works well in most real-life environments, even the most challenging ones, like non-stationary background noise. In the presence of a babble background noise with an SNR of 0dB the rate of undetected words is about 3.5%.