Speech endpoint detection is one of the key problems in the practical application of speech recognition system. In this paper, speech signal contained chirp is decomposed into several intrinsic mode function (IMF) with the method of ensemble empirical mode decomposition (EEMD). At the same time, it eliminates the modal mix superposition phenomenon which usually comes out in processing speech signal with the algorithm of empirical mode decomposition (EMD). After that, selects IMFs contained major noise through the adaptive algorithm. Finally, the IMFs and speech signal contained chirp are input into the independent component analysis (ICA) and pure voice signal is separated out. The accuracy of speech endpoint detection can be improved in this way. The result shows that the new speech endpoint detection method proposed above is effective, and has strong anti-noises ability, especially suitable for the speech endpoint detection in low SNR.