Voice activity detection (VAD) is more and more essential in the noisy environments to provide an accuracy performance in the speech recognition. In this paper, we provide a method based on left-right hidden Markov model (HMM) to identify the start and end of the speech. The method builds two models of non-speech and speech instead of existed two states, formally, each model could include several states, we also analysis other features, such as pitch index, pitch magnitude and fractal dimension of speech and non-speech.. We compare the VAD results with the proposed algorithm and two states HMM. Experiments show that the proposed method make a better performance than two states HMMs in VAD, especially in the low signal-to-noise ratio (SNR) environment.