“A source-filter separation for non-stationary voiced speech based on sinusoidal representation