AM-FM decomposition of speech signals: An asymptotically exact approach based on the iterated Hilbert transform

This paper presents a multicomponent sinusoidal model of speech signals, obtained through a rigorous mathematical formulation that ensures an asymptotically exact reconstruction of these nonstationary signals, despite the presence of transients, voiced segments, or unvoiced segments. This result has been obtained by means of the iterated use of the Hilbert transform, and the convergence properties of the proposed method have been both analytically investigated and empirically tested. Finally, an adaptive segmentation algorithm used to accurately compute instantaneous frequencies from unwrapped phases, suited to complete the proposed AM-FM model, is presented.