Stochastic modeling for automatic speech understanding