BACKGROUND: Electromyography is a non-stationary and non-Gaussian physiological signal. Currently, the high-order spectral technique, which based on higher-order cumulant, has been widely used in solving problems such as non-Gaussian, non-stationary and nonlinearity. OBJECTIVE: To propose a feature extraction method for surface electromyography (SEMG) signals based on a non-Gaussian AR model parameterized bispectrum estimation and fisher linear discriminant analysis. METHODS: Aim at features of SEMG, from point of high statistics, and based on a non-Gaussian AR model, bispectrum analysis was performed to extract effective features, followed by constructing characteristic vector by fisher linear discriminant analysis dimension reduction, then the support vector machine was used to classify the movement patterns. The differences of recognition rates between AR+BIS+LDA and other features extracted by different methods were compared. RESULTS AND CONCLUSION: Experimental results showed that the eight forearm movement patterns could be well identified after training by multi-class support vector machine and its average recognition rate reached above 97.6%. For short data sets, bispectrum's feature had a better pattern recognition rate than other features such as AR model coefficients, wavelet packet transformation coefficients. That improved the performance of real-time control of prosthesis.