Abstract: Combined Empirical Mode Decomposition（EMD） of Hilbert-Huang Transform（HHT） with Recurrence Plot（RP） method，a new algorithm of speech endpoint detection is proposed.In this algorithm，different Intrinsic Mode Function（IMF） is filtered on the time scale of soft threshold by EMD and multi-scale features based on speech and noise at first.Then，through using the RP method of nonlinear dynamic behavior，quantitative statistical analysis of the deterministic recursive to carry out speech endpoint detection.Simulation results show that the method has strong ability of non-steady-state analysis of dynamic change，more accurately than the traditional methods to extract the start and end point of speech signal and better robustness in low SNR environment.