A new method for determining the vocal tract transfer function and its excitation from voiced speech

A method for determining both the transfer function and the excitation of the vocal tract from voiced speech is described. The method involves estimating the parameters of an autoregressive (AR) speech model containing an explicit representation of the voice source. The speech model is parameterized in terms of the AR parameters, the instants at which the glottis opens and closes, and the derivatives of the glottal pulse at these instants. Although linear in the AR parameters and the glottal pulse derivatives, the model is nonlinear in the glottal endpoints. The estimation procedure involves carrying out a direct search of the glottal endpoints, and at each point in the search minimizing the total squared model error with respect to the linear parameters. Experimental results from real vowels show that the method gives prediction gains that are between 3 and 10 dB better than using a linear predictive coding