Abstract: Aiming at the fact that the existing speech intelligibility evaluation methods cannot effectively deal with the signal changes under various types of nonlinear distortions, a bispectral speech intelligibility metric (BSIM) algorithm based on bispectral features is proposed, which uses third-order statistics to extract features from the spectrogram of speech signal. Bispectrum can detect the nonlinear phase coupling in the speech signal and suppress the Gussian noise in the non-Gussian signal, thereby can reveal more useful information hidden in the signal. This method is compared with existing speech intelligibility indicators. The results show that this method can successfully predict the degradation of speech intelligibility caused by linear distortion and nonlinear distortion. The evaluation result is highly correlated with the subjective intelligibility result and sensitive to signal distortion changes.