Abstract: Empirical mode decomposition (EMD) is a local, fully data driven and self-adaptive analysis approach. It is a powerful tool for analyzing multi-component signals. Aiming at the reduction of scale mixing and artificial frequency components, an improved scheme of EMD is proposed for the analysis and reconstruction of nonstationary and multicomponent speech signals. The improved EMD method uses higher order extrema and then the optimal envelope mean can be obtained by an inverse EMD scheme. A new sifting stop criterion is proposed based on the index of orthogonality. And finally, the mean square error criterion is used for performance evaluation. Experiments on simulation signals and actual sound signals prove the correctness and validity of the modified method. An exact decomposition result can be obtained by the improved EMD method with less scale mixing than the standard EMD method. The proposed method has broad application potential in processing of speech signals, mechanical vibration signals and radar signals, etc.