Stochastic glottal source applied to voiced-speech decomposition using state-space methods

Glottal source estimation is a difficult task in voiced-speech signal processing. In the past, several deterministic models of the glottal source have been proposed. However, those models can not suitable represent the perturbations or the aperiodicities occurring in real voices. In this work, an alternative glottal source model is formulated, ruled by stochastic difference equations and inspired by Liljencrants-Fant glottal function. In agreement with the voice source-filter theory, a state-space model for voiced-speech signals is also implemented, combining the stochastic glottal source with a linear time-varying vocal tract filter. These two models are applied to pitch-synchronous voice decomposition into glottal source and vocal tract components using state-space methods. Simulations with synthesized signals are exposed, showing that the estimates of the glottal source and of the vocal tract can be suitably calculated for all the considered conditions. Moreover, preliminary results are presented, suggesting that the proposed method could also be useful for real voice decomposition.