In order to improve the accuracy of voice activity detection in car environment, we propose a neural network structure based on gated recurrent unit-recurrent neural network (GRU-RNN) to process the Log-Mel feature sequence of noisy speech and separate speech from noise. After restoring the Log-Mel sequence of clean speech, a new feature, Log-Mel-Sum, is proposed to detect voice activity. Experimental results show that the proposed method has a high accuracy of voice activity detection.