In this paper we introduce an efficient genetic algorithm based voice activity detection (GA-VAD) algorithm. The inputs for GA-VAD are zero-crossing difference and a new feature that is extracted from signal envelope parameter, called MULSE (multiplication of upper and lower signal envelope). The voice activity decision is obtained using a Threshold algorithm with additional decision smoothing. The key advantage of this method is its simple implementation and its low computational complexity and introducing a new simple and efficient feature, MULSE, for solving the VAD problem. The MULSE parameter could be appropriate substitution for energy parameter in VAD problems. The GA-based VAD algorithm (GA-VAD) is evaluated using the Timit database. It is shown that the GA-VAD achieves better performance than G. 729 Annex B at any noise level with a high artificial-to-intelligence ratio.