The sign language is a complex motion pattern with dynamic changes in various gestures. The effect of gesture feature processing is directly related to the accuracy of sign language recognition. In this article, a new dynamic gesture feature processing method based on the estimation of improved S-transform (IST) spectral of surface electromyography (sEMG) is proposed. The collected sEMG signal is transformed by S-transform, the optimization factor is introduced to adjust the time-frequency resolution, and the IST spectrum is generated. The time and frequency components of the IST spectrum are defined as 2-D random variables, and the matrix elements of the IST spectrum are taken as 2-D random variables. The 2-D kernel density function is obtained by Gaussian kernel density estimation. Simulation and experiments show that the estimation method of the IST spectrum effectively suppresses white noise and strengthens the sEMG transient mutation characteristics of dynamic gestures. Compared with empirical mode decomposition, selfpermutation entropy, and singular value permutation entropy, the accuracy of dynamic gesture recognition based on this method is improved by 10. 0% , 6. 67% and 11. 67% , respectively.