The paper compares a sliding-mode observer with the standard and extended versions of the Kalman filter for full state estimation in an induction machine. The design method for the sliding-mode observer is presented and shown to possess invariant dynamic modes, which can be assigned independently to achieve the desired performance. The chattering concerned with the sliding-mode method can be eliminated by modifying the observer gain, without sacrificing robustness and precision. The indices of comparison are dynamic performance, robustness against parameter uncertainties, noise sensitivity, stability and computational complexity. Simulation and experimental results are presented in this paper. Important conclusions, together with recommendations for observer selection, regarding the effectiveness of the various schemes are given.