Logarithmic Energy (LE) and Spectral Entropy (SE) were integrated to form a new characteristic, Logarithmic Energy Spectral Entropy (LESE), by using Fuzzy C Means Clustering algorithm and Bayesian Information Criterion algorithm to estimate the thresholds of the LESE characteristic, and by using dual threshold method for voice activity detection. Experiments on the TIMIT continuous speech database have shown that, compared with the Energy spectral Entropy (EE) and LE in the noisy environments, the LESE has better performance not only in detection but also in robustness. When the SNR is -5 dB, the detection error rate of the LESE is only 18.02%, and while the SNR is 0~10 dB, its detection error rate is significantly lower than the EE and LE.