A new method for automatically fitting the Liljencrants-Fant (LF) model to the time domain waveform of the glottal flow derivative is presented in this paper. By applying an extended Kalman filter (EKF) to track the LF-model shape-controlling parameters and dynamically searching for a globally minimal fitting error, the algorithm can accurately fit the LF-model to the inverse filtered glottal flow derivative. Experimental results show that the method has better performance for both synthetic and real speech signals compared to a standard time-domain LF-model fitting algorithm. By offering a new method to estimate the glottal source LF-model parameters, the proposed algorithm can be utilised in many applications.