Abstract: A Nonlinear Weighting Energy feature,which is used to investigate the differences of pitch-synchronous dynamic frame-length energies among phonemes in each words is proposed,in order to solve the robustlessness problem of traditional features in lexical stress detection.The contribution of Nonlinear Weighting Energy feature to English lexical stress detection was evaluated with ISLE database.Experimental results show that the Nonlinear Weighting Energy feature is more robust than traditional features,while the combination of Nonlinear Weighting Energy and traditional features could provide a reduction of 3.58% in terms of error rates compared with the results using traditional features only.