This study deals with estimation of the mechanical twitch of a single motor unit in human muscle during voluntary contraction. The existing estimation method is based on averaging of the force produced by the whole muscle, using the motor unit action potential for triggering. This method leads to underestimation of twitch amplitude and rise time due to partial fusion of the twitches, even at the lowest firing rates which can be maintained during voluntary contraction. To obtain unbiased twitch estimates, even when the twitches partially fuse, three versions of a nonparametric system identification method are explored: averaging plus least squares estimation; least squares estimation; and Markov estimation. Averaging plus least squares estimation give poorer noise reduction than least squares estimation. The noise reduction obtained by least squares estimation is about the same as the noise reduction obtained by averaging alone. The noise is mainly due to the contraction of all active motor units, except the unit being studied. This noise has dominating low-frequency components. Markov estimation takes the spectrum of the noise into account, and thus in most cases provides a better estimate than least squares estimation.