In this paper, a robust voice activity detection algorithm based on a long-term metric using dominant frequency and spectral flatness measure is proposed. The propose algorithm makes use of the discriminating power of both features to derive the decision rule. This method reduces the average number of speech detection errors. We evaluate its performance using 15 additive noises at different SNRs (-10 dB to 10 dB) and compared with some of the most recent standard algorithms. Experiments show that our propose algorithm achieves the best performance in terms of accuracy rate average over all SNRs and noises.