Estimation of Speech Features of Glottal Excitation by Nonlinear Prediction

Analysis of speech signals can be performed with the aid of linear or nonlinear statistics using appropriate prediction algorithms. In this contribution, speech features are treated using the results of a nonlinear prediction based on Volterra series. Features are investigated representing the prediction gain by nonlinear statistics and representing individual coefficients of the nonlinear components. The features are estimated quasi continuously resulting in a feature signal. Additionally, to obtain features which are highly sensitive to segmentation shifting, an asymmetric window function is integrated into the prediction algorithm. The analyses of speech signals show that the estimated features correlate with the glottal pulses. Furthermore, the investigations show that using the first individual nonlinear coefficient as a feature is advantageous over using the prediction gain.