In this paper we develop a voice activity detection algorithm based on spectrum estimation of speech and non-speech segments using Vector Quantization method. In this method, we try to classify entry speech signal to speech and non-speech classes. Commonly, the performance of the voice activity detection (VAD) algorithms in non-stationary background noise is not so satisfying under low SNR, so we try to concentrate our study on this issue. The model of a non-speech is a codebook generated from noise and model of speech is several codebook generated from speech contaminated by noise in some different SNR. The labeling is performed by evaluating the distortions between the entry signal samples and the designed models. Our simulation results based on the Persian speech database show that the VQ based VAD is high performance in low SNR conditions (SNR < 5 dB).