A stochastic model for the speech sonority