LF model based glottal source parameter estimation by extended Kalman filtering

A new algorithm for glottal source parameter estimation of voiced speech based on the Liljencrants-Fant (LF) model is presented in this work. Each pitch period of the inverse filtered glottal flow derivative is divided into two phases according to the glottal closing instant and an extended Kalman filter is iteratively applied to estimate the shape controlling parameters for both phases. By searching the minimal mean square error between the reconstructed LF pulse and the original signal, an optimal set of estimates can be obtained. Preliminary experimental results show that the proposed algorithm is effective for a wide range of LF parameters for different voice qualities with different noise levels, and accuracy especially for estimation of return phase parameters compares better than standard time-domain fitting methods while requiring a significantly lower computational load