A nonlinear Hammerstein model is proposed for coding speech signals. Using Tsay's nonlinearity test, we first show that the great majority of speech frames contain nonlinearities (over 80% in our test data) when using 20-millisecond speech frames. Frame length correlates with the level of nonlinearity: the longer the frames the higher the percentage of nonlinear frames. Motivated by this result, we present a nonlinear structure using a frame-by-frame adaptive identification of the Hammerstein model parameters for speech coding. Finally, the proposed structure is compared with the LPC coding scheme for three phonemes /a/, /s/, and /k/ by calculating the Akaike information criterion of the corresponding residual signals. The tests show clearly that the residual of the nonlinear model presented in this paper contains significantly less information compared to that of the LPC scheme. The presented method is a potential tool to shape the residual signal in an encode-efficient form in speech coding.