Novel stochastic properties of the short-time spectrum for unvoiced pronunciation modeling and synthesis.