In this paper, a fractal dimension algorithm is proposed based on the neighborhood extremum difference signal and its power spectrum. The proposed method is applied to speech activity detection (SAD)in low SNR environments. In the time domain, the extremum difference signal is searched in the neighborhood.The fractal value is then estimated from the power spectrum of the difference signal based on a minimum error criterion. In a quiet environment, performance of the method is similar to the box algorithm and better than entropy algorithm in normal and whispered speech detection, while in several noise environments, it clearly outperforms the entropy algorithm. It is also better than the box algorithm except in a white noise
environment. In addition, the computation load is only 5% of the box algorithm. Experimental results show that the proposed algorithm has a good overall performance in terms of efficiency and SAD.