Gaussian-weighted voice inverse filtering: Effects of varying the attenuation window parameters on the glottal source estimation

IVÁN ARIEL ZALAZAR; GABRIEL ALEJANDRO ALZAMENDI; GASTÓN SCHLOTTHAUER

San Juan

Congreso; XIX Reunión de Trabajo en Procesamiento de la Información y Control, RPIC 2021; 2021

Instituto de Automática, Facultad de Ingeniería, Universidad Nacional de San Juan

Inverse filtering allows to indirectly obtain estimates of the glottal source which carry relevant information of the vocal function. Inverse filtering strategies are generally based on variations of the linear prediction model for the phonation time series. Unfortunately, the classical formulation of linear prediction is very susceptible to outliers such as those occurring during maximum excitation instant in the voice signal. Hence, the weighted linear prediction becomes a valuable alternative, where a weight function is applied to attenuate the relative effect of the voice samples around the maximum excitation instants. In this article, a weighted linear prediction using a Gaussian weight function is revisited. A new parameterization for the Gaussian attenuation involving the fundamental frequency is introduced. Then, the effects of the different parameters of the weight function on the glottal source estimation are analyzed, using synthesized voice signals and real examples. Our results for the synthesized signals indicate that optimal ranges for the weight function parameters can be obtained that results in a minimal glottal source estimation error. Evidence is also provided that the weighted inverse filtering with Gaussian attenuation is also suitable for the real voice signals whenever the optimal parameters are considered.