A novel voiced speech enhancement approach based on modulated periodic signal extraction

Most of the existing speech coding and speech enhancement techniques are based on the AR model and hence apply well to unvoiced speech. These same techniques are then applied to the voiced case as well by extrapolation. However, voiced speech is very structured so that a proper approach allows to go further than for unvoiced speech. We model a voiced speech segment as a periodic signal with (slow) global variation of amplitude and frequency (limited time warping). The bandlimited variation of global amplitude and frequency gets expressed through a subsampled representation and parameterization of the corresponding signals. Assuming additive white Gaussian noise, a Maximum Likelihood approach is proposed for the estimation of the model parameters and the optimization is performed in an iterative (cyclic) fashion that leads to a sequence of simple least-squares problems. Particular attention is paid to the estimation of the basic periodic signal, which can have a non-integer period, and the estimation of the amplitude signal with guaranteed positivity.