under strong noise environments, the speech detection often performs bad, in order to make some improvements the Hilbert-Huang Transform is used in the algorithm. The speech signal is decomposed into finite Intrinsic Mode Functions, and then, with the Hilbert transform, the energy-frequency-time distribution of the original signal can be obtained. The EMD is used as a filter to remove unwanted noise, and then the feature was extracted to detect speech frames by investigating the distribution of energy depending on the time. Experiments show HHT is helpful to extract the characteristic parameters of the signals, and also is capable to improve the performance of speech detection.